← Network
Model Card v3.1.2
Fidelity-based behavioral measurement for network traffic.
Version
v3.1.2
CICIDS Fri AUC
0.9990
MAWI 2021
0.924
Perspectives
4
CPU only
Air-gapped
Download model card Download benchmark PDF Run live demo

Accuracy Profiles

All results: 5-fold cross-validated. Calibration on benign only. Default α=0.05, with a tighter operating point (α=0.005) shown alongside where relevant.

CICIDS-2017 (Enterprise, multi-scale)

DayFlowsAttack %AUC±F1 @ α=0.05F1 @ α=0.005
Friday500,00046.9%0.99900.0000.9720.992
Wednesday496,64135.7%0.99860.0000.9570.995
Tuesday322,0782.2%0.99850.0000.4700.898
Thursday362,07620.4%0.99840.0000.9100.969

Full operating points at α=0.05: Friday P 0.946 R 0.999, Wednesday P 0.918 R 1.000, Tuesday P 0.307 R 1.000, Thursday P 0.836 R 0.997. FPR is 5.0% by construction at this α.

Tuesday shows what the α parameter does. At 2.2% attack prevalence, a 5% false positive budget admits more benign false positives than there are attacks to find, so F1 is 0.470 while AUC is 0.9985. Moving α to 0.005 recovers F1 to 0.898 with no change to the instrument.

Mode-aware calibration provides measurable improvement on heterogeneous traffic and is neutral to slightly negative where the benign population is already homogeneous. Reported as a conditional gain, not a uniform one.

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

Random-sampled 500K flows from full daily captures (5.4M to 7.4M rows each).

DayFlowsAUC±F1Attack Types
Fri 16-02500,0001.0000.0000.948DoS Hulk (1.8M in full capture)
Tue 20-02500,0001.0000.0000.773DDoS-LOIC (289K in full capture)
Wed 21-02500,0001.0000.0000.929DDoS-HOIC (1.08M in full capture)
Wed 14-02500,0000.9930.0000.781FTP/SSH Brute Force
Thu 15-02500,0000.9930.0000.348DoS GoldenEye/Slowloris
Fri 02-03500,0000.9940.0000.592Botnet Ares
Wed 28-02500,0000.9950.0000.318Infiltration
Thu 01-03500,0000.9910.0000.287Infiltration/NMAP
Fri 23-02500,0000.9970.0010.003Web Attack (20 flows in sample)
Thu 22-02500,0000.9980.0000.002Web Attack (18 flows in sample)

Three AUC=1.000 days are volumetric floods, expected. Two F1≈0 days have <25 attacks in 500K, base rate. Attack-weighted F1 across all 10 days: 0.876 (versus supervised F1 0.744).

MAWIFlow (Real Backbone)

YearFlowsMethodAUC±F1 @ α=0.05F1 @ peak α
2016†500,000Flow-only0.9540.0010.6080.634 (α=0.07)
2021†500,000Multi-scale0.9240.0010.5210.547 (α=0.07)
2021†500,000Flow-only0.8600.0070.5170.525 (α=0.07)

† Random sample. Multi-scale uses the full measurement stack; flow-only uses the per-flow measurement alone.

Both methods are shown for 2021 because they differ materially: multi-scale ranks better (0.924 against 0.860) while flow-only reaches a similar F1 at the operating point. Each year is reported under the method that is stable for it. Multi-scale on 2016 is not stable and is not reported; multi-scale on 2021 is independently reproduced.

MAWI is the one benchmark family here drawn from real backbone traffic rather than a laboratory capture, and it is the family where AUC and the achievable operating point diverge most. Ranking on 2016 is strong at 0.954 while peak F1 is 0.634. Both are reported because the gap is the honest description of how the instrument behaves on production traffic. Do not tighten α on this traffic: at α=0.005 MAWI 2016 F1 falls to 0.150 and MAWI 2021 to 0.118.

MAWI 2011 (Real Backbone, multi-scale)

YearFlowsAUC±
201176,3810.8920.002

Independently reproduced at 0.8919.

Other Domains (flow-level)

DatasetYearDomainFlowsAUC±F1PRFPR
NSL-KDD Train1999Legacy125,9730.9930.0010.9530.9440.9712.5%
UNSW Train2015Modern82,3320.9930.0020.9610.9600.9772.3%
UNSW Test2015Modern175,3410.9800.0010.9240.9750.9142.7%
NSL-KDD Test1999Legacy22,5430.9690.0020.8780.9600.9122.6%
TON-IoT2021IoT500,0000.9480.0010.6460.7970.5523.6%
DoH2020Encrypted269,6430.9380.0010.6810.9920.5195.0%
IoT-DIAD2024IoT500,0000.8640.0030.5040.7790.6913.0%

BCCC-IoT (2024, smart home) is withheld from this release pending end-to-end reproduction of a previously published figure.

DoH: 0.931 to 0.938 after a feature mapping correction. Precision ≥ 0.989 at all α.

GPU Cluster Behavioral Fidelity (same measurement principles)

BenchmarkMetricResult
GWDG A100 GPU failuresDetection16/16 (100%)
GWDG healthy baselinesFalse alarm0/5 (0%)
Lingjun labeled GPU hostsAUC (5-fold)0.952 ± 0.007
Lingjun labeled GPU hostsF1 at α=0.07~0.80
Speed (Mac Mini)Throughput21,000+ flows/sec

Comparison to Supervised Baselines

BenchmarkVERITY (no labels)SupervisedMethod
MAWI 2016 (flow-only)AUC 0.954AUC 0.902RF (Schraven et al. 2026)
MAWI 2021 (multi-scale)AUC 0.924AUC 0.903RF (Schraven et al. 2026)
MAWI 2021 (flow-only)AUC 0.860AUC 0.903RF (Schraven et al. 2026)
CIC-2018 (weighted F1)F1 0.876F1 0.744RF (arXiv 2606.29797)
CICIDS-2017 FriAUC 0.9990AUC 0.999CNN-BiLSTM
NSL-KDD TrainAUC 0.993AUC 0.997XGBoost

On MAWI, label-free measurement beats the supervised baseline in 2016 by 0.052 and in 2021 by 0.021 under multi-scale. Flow-only in 2021 trails the baseline by 0.043. Both methods are shown so the margin is not overstated.

---

Cross-Temporal Resilience

TransferAUCNotes
5-year forward (2011 → 2016)0.942Ranking holds; threshold must be re-derived
10-year forward (2011 → 2021)0.531Comparable to supervised RF at 0.61
Same-year 20160.952Cross-year feature basis
Same-year 20210.853Cross-year feature basis

What transfers is the ranking, not the threshold. Thresholds must be re-derived on the benign traffic of the period being defended. That re-derivation is the ordinary recalibration path and costs minutes of benign observation.

Against the supervised comparator, ten-year transfer degrades for both approaches. The operational difference is recovery cost: recalibrate on CPU in minutes versus relabel and retrain on GPU over weeks.

---

Operating Point Guidance

α is the one operator parameter. The shipped default of 0.05 is not the best value on every dataset. Set α from how cleanly the instrument separates the traffic on a representative calibration sample: tighten when separation is strong, leave looser when it is weak. Tighten further when attack prevalence is low. Do not tighten on weakly separated traffic; a tight budget there starves recall.

Changing α changes no model parameter and costs nothing at runtime.

---

FPR Consistency

At α=0.05, observed FPR ranges 2.3% to 5.3% across all evaluated datasets (1999 to 2024). FPR is controlled by α.

---

Documented Limitations

Calibration assumes a clean benign observation window. Contamination checks warn the operator; automated integrity verification is still shipping.

Cross-temporal capability is ranking transfer with mandatory threshold re-derivation, not end-to-end detection transfer. Decade-scale transfer also degrades in ranking (AUC 0.531, against supervised RF 0.61).

Mode-aware calibration is a conditional gain, not a uniform one. It helps on heterogeneous benign populations and is slightly negative on homogeneous ones.

The default α understates achievable F1 on well-separated traffic. See Operating Point Guidance.

The encoder is trained on enterprise and academic benign traffic, so IoT and encrypted DNS domains are weaker (DoH AUC 0.938 after a feature mapping correction). Adversarial evasion testing against adaptive mimicry has not been conducted.

Multi-scale evaluation is not stable on every backbone capture. Where it is unstable, results are reported flow-only. Where it is stable, both methods are shown. Instability is dataset-specific rather than a general property of backbone traffic.

---

Credasis · Patent Pending

For customers, investors, and evaluators. Methodology for peer review is in the companion technical paper.