VERITY Network Intelligence — Model Card v2.1. VERITY measures whether network traffic behavior is consistent with its calibrated baseline across multiple independent dimensions simultaneously. When behavior deviates, the deviation is the detection.
No attack labels are used for calibration or detection. Labels are used only for benchmark scoring, fold construction, and benign/attack evaluation bookkeeping. No GPU. No internet. Calibrates from benign traffic observation. Every number in this document is reproducible from the deployment package on publicly available datasets.
VERITY v2.1 ships two operating modes from a single codebase.
The operator selects mode and sensitivity level. Everything else is calibration-emergent from the deployment environment's own benign traffic.
| Study | Dataset | Purpose | Result |
|---|---|---|---|
| Primary benchmark | CICIDS-2017 (Engelen 2021) | Per-class detection, 15 classes | 14/15 ≥ 93% (Detection) |
| Cross-dataset | CSE-CIC-IDS2018 (Distrinet) | Generalization without retuning | Precision mode validated; Detection mode artifact pending |
| Operating curves | CICIDS-2017, 4 days | Operator sensitivity tradeoff | Full α ∈ [0.001, 0.10] |
| Temporal stability | CICIDS-2017 (60/40 split) | Detection under time shift | F1=0.946 Friday |
| Calibration poisoning | CICIDS-2017 | Resilience to contamination | Detected at 1%, refused |
| Adversarial evasion | CICIDS-2017 Tuesday | Spacing/jitter on Patator-class attacks | Complementary coverage on tested scenarios |
| Head ablation | CICIDS-2017 Friday | Per-dimension importance | Critical dimension identified |
| Production E2E | CICIDS-2017 Tuesday | Guard-enabled full chain (Patator) | FTP-Patator R=0.995, checksummed |
| Additional datasets | IoT, Darknet, HIKARI, DoH | Domain breadth | Varies by domain |
All CICIDS-2017 numbers use the Engelen et al. 2021 corrected release — not the widely circulated UNB TrafficLabelling/ CSVs, which contain mislabeled and unlabeled attack flows. Download corrected files from the live demo.
5-fold cross-validated. Engelen 2021 corrected. Clean threshold protocol (training benign only).
| Class | n | Precision Mode | Detection Mode | Improvement |
|---|---|---|---|---|
| DoS Hulk | 158,468 | 1.000 | 1.000 | — |
| DDoS | 95,144 | 1.000 | 0.999 | — |
| Botnet | 736 | 1.000 | 1.000 | — |
| Web Brute Force | 73 | 1.000 | 1.000 | — |
| Web XSS | 18 | 1.000 | 1.000 | — |
| Heartbleed | 11 | 1.000 | 1.000 | — |
| DoS Slowhttptest | 1,740 | 0.998 | 0.998 | — |
| Infiltration-Portscan | 71,767 | 0.993 | 0.998 | +0.005 |
| SSH-Patator | 2,961 | 0.988 | 0.995 | +0.007 |
| Infiltration | 36 | 0.980 | 1.000 | +0.020 |
| Portscan | 159,066 | 0.947 | 0.952 | +0.005 |
| DoS GoldenEye | 7,567 | 0.936 | 0.939 | +0.003 |
| DoS Slowloris | 3,859 | 0.912 | 0.984 | +0.072 |
| FTP-Patator | 3,972 | 0.603 | 0.995 | +0.392 |
| SQL Injection | 13 | 0.067 | 0.077 | +0.010 |
| Summary | Precision | Detection |
|---|---|---|
| Classes ≥ 93% recall | 11 of 15 | 14 of 15 |
| F1 (Friday) | 0.963 | 0.951 |
| FPR | 3.8% | 6.3% |
Detection mode gains three class upgrades (FTP-Patator, Slowloris, Infiltration) at the cost of 2.5% additional FPR.
| α | Precision | Recall | FPR | F1 |
|---|---|---|---|---|
| 0.001 | 0.996 | 0.283 | 0.1% | 0.440 |
| 0.005 | 0.990 | 0.478 | 0.4% | 0.644 |
| 0.01 | 0.984 | 0.650 | 0.8% | 0.784 |
| 0.02 | 0.973 | 0.836 | 1.5% | 0.902 |
| 0.03 | 0.964 | 0.944 | 2.3% | 0.954 |
| 0.05 | 0.959 | 0.968 | 3.8% | 0.963 |
| 0.07 | 0.945 | 0.976 | 5.1% | 0.960 |
| 0.10 | 0.927 | 0.993 | 7.1% | 0.959 |
| α | Precision | Recall | FPR | F1 |
|---|---|---|---|---|
| 0.001 | 0.997 | 0.626 | 0.1% | 0.765 |
| 0.005 | 0.991 | 0.905 | 0.5% | 0.946 |
| 0.01 | 0.984 | 0.947 | 0.9% | 0.965 |
| 0.02 | 0.971 | 0.970 | 1.6% | 0.971 |
| 0.03 | 0.958 | 0.976 | 2.4% | 0.967 |
| 0.05 | 0.936 | 0.991 | 3.8% | 0.963 |
| 0.10 | 0.888 | 1.000 | 7.0% | 0.940 |
| Day | Prec F1 | Prec R | Det F1 | Det R |
|---|---|---|---|---|
| Wednesday (DoS) | 0.963 | 0.991 | 0.944 | 0.997 |
| Friday (DDoS/Portscan) | 0.963 | 0.968 | 0.951 | 0.970 |
| Thursday (Infiltration) | 0.915 | 0.967 | 0.874 | 0.972 |
| Tuesday (Patator) | 0.431 | 0.603 | 0.403 | 0.995 |
Tuesday F1 is dominated by the benign/attack ratio (315K benign, 7K attacks). Per-class recall is the operationally meaningful metric.
Same engine. Same parameters. No modification for CIC-2018.
| File | Key class | n | Precision R | Detection R |
|---|---|---|---|---|
| Wed-21 | DDoS-HOIC | 248,069 | 1.000 | 1.000 |
| Wed-14 | SSH-BruteForce | 94,197 | 1.000 | 1.000 |
| Tue-20 | DDoS-LOIC-HTTP | 246,049 | 0.980 | 1.000 |
| Thu-01 | Inf-NMAP Portscan | 17,407 | 0.990 | 0.999 |
| Fri-23 | Web XSS / Brute Force | 73 / 62 | 1.000 | 1.000 |
| Dataset | Domain | Best F1 | at α | Key finding |
|---|---|---|---|---|
| BCCC-IoT | IoT security | 0.887 | 0.10 | 10 IoT attack classes |
| Darknet | Encrypted traffic | 0.295 | 0.10 | Tor R=0.859, VPN R=0.146 |
| HIKARI 2022 | Bruteforce | 0.602 | 0.02 | Bruteforce R=1.000 |
| CIC-DoH | Encrypted DNS | 0.716 | 0.10 | Malicious R=0.561 |
| Contamination | Without guard | With guard |
|---|---|---|
| 0% | F1=0.948 | Deploy |
| 1% | F1=0.663 (collapse) | Refuse — contamination detected |
| 3% | F1=0.297 | Refuse |
| 5% | F1=0.164 | Refuse |
| Property | Precision mode | Detection mode |
|---|---|---|
| Engine size | 500 KB | 560 KB |
| Total deployment | ~50 MB | ~70 MB |
| RAM at runtime | ~200 MB | ~250 MB |
| Per-flow latency | ~1 ms | ~4 ms |
| Throughput | ~1M flows/sec | ~250K flows/sec |
| Calibration time | Under 30 seconds | Several minutes |
| Hardware | CPU only | CPU only |
| Labels required | None | None |
| Internet required | No | No |
| Air-gapped deployment | Yes | Yes |
| Limitation | Detail |
|---|---|
| SQL Injection (n=13) | R=0.077. Payload-level attack, below flow metadata resolution. |
| Detection mode FPR | 6.3–7.6%. Cost of detecting geometric-anomaly attacks. |
| Calibration poisoning | Catastrophic without guard. Detected and refused with guard. |
| Adversarial evasion | Tested on feature jitter and density reduction (Patator class). Low-and-slow and protocol-aware mimicry untested. |
CICIDS-2017: Canadian Institute for Cybersecurity, University of New Brunswick. Sharafaldin, Lashkari & Ghorbani (ICISSP 2018).
Improved CICIDS-2017: Engelen, Rimmer & Joosen (IEEE SPW 2021). Distrinet Research Group, KU Leuven.
CSE-CIC-IDS2018: Communications Security Establishment & Canadian Institute for Cybersecurity, with Distrinet/KU Leuven corrections.
VERITY Network Intelligence v2.1 — Model Card — June 2026 — Credasis AI Inc. — Patent Pending