Network
The attack does not need a name to be detected.

VERITY Network Intelligence v2.1 is zero-day anomaly detection by design. No signatures. No training on attack data. No cloud dependency. Calibrate on benign traffic only — attack labels are not used for calibration or detection. The deviation is the detection.

0.963
Friday F1 (Precision)
14/15
Classes ≥ 93% Recall
Zero
Attack Labels Used
Two modes
Precision & Detection
Model card v2.1 →
What it is

Detection without prior knowledge.

"Signatures catch what has been seen. Behavioral measurement catches what has not. The attack that has never been catalogued is still a deviation from the baseline — and deviation is measurable."

Signature-based detection works by pattern matching. It is fast, precise, and structurally limited to attacks that have been seen before. Zero-day attacks pass through invisible. So do the long tail of attacks that were never important enough to receive their own signature. The industry response has been to wait for the breach, write the signature, and hope no one else gets hit first.

VERITY takes a different approach. It calibrates on the benign traffic in your environment. It learns the behavioral shape of normal flows — volume, timing, directionality, session structure. Every incoming flow is measured against that baseline. The attack does not need a name. It does not need a signature. It does not need to have been seen before. It only needs to be different from normal, and difference is measurable.

What it does

The deviation is the detection.

VERITY calibrates on your benign traffic in a single pass. Every flow is then measured against multiple independent geometric reference points. Flows that deviate are flagged. Flows near the decision boundary are escalated. When the system is uncertain, it says so.

Signature-based detection
Flow detected. Signature database queried. No match. Classification: benign. Action: none.

The attack is novel. There is no signature for it yet. The detector cannot distinguish it from legitimate traffic. The flow passes through.
VERITY measures
Flow exhibits behavioral characteristics that deviate from calibrated baseline. Multiple independent measurements converge on the deviation. Forensic signature generated. Verdict: anomalous. Decomposition: byte volume +6.8σ, session duration +3.2σ, packet rate variance +4.1σ.

The system has never seen this attack. It does not need to.
01
Calibrate on normal
Point VERITY at your benign traffic. In a single pass it learns the behavioral shape of your network — volume, timing, directionality, uniformity, session patterns. No attack data required. No labels. No training.
02
Measure every flow
Each incoming flow is measured against the calibrated baseline using multiple independent geometric tests. The engine requires convergence across measurements before committing to a verdict. A single noisy signal cannot drive a false positive.
03
Explain every flag
Every flagged flow includes a decomposition of which behavioral dimensions deviated and by how much. Not an opaque anomaly score. A specific, verifiable measurement. Every alert carries a 28-byte forensic signature for deduplication and review.
Validated

Engelen-corrected CICIDS-2017 benchmarks.
Precision and Detection modes.

5-fold cross-validation on the Engelen 2021 corrected CICIDS-2017 release — not the widely circulated mislabeled UNB CSVs. Clean threshold protocol: training benign only. Two operating modes from a single v2.1 codebase. We report every class, including weak ones like SQL Injection at flow-metadata resolution.

DatasetTypeModeF1RecallNotes
CICIDS-2017 FridayDDoS / PortScan / BotnetPrecision0.9630.968α=0.05, Engelen corrected
CICIDS-2017 FridayDDoS / PortScan / BotnetDetection0.9510.970Genesis S2 rescue enabled
CICIDS-2017 WednesdayDoS variantsPrecision0.9710.970α=0.02
CICIDS-2017 TuesdayFTP / SSH PatatorDetection0.995FTP-Patator recall at α=0.05
CSE-CIC-IDS2018Cross-datasetPrecision0.9961.000No retuning
CSE-CIC-IDS2018Cross-datasetDetection1.000LOIC-HTTP, HOIC, Hulk improved

14 of 15 attack classes at or above 0.93 recall in Detection mode. SQL Injection (n=13) remains weak — payload-level, below flow-metadata resolution. Optional corrected benchmark CSVs on the live demo.

Full benchmark results → Download PDF Model card v2.1 →
How it compares
PropertyVERITYSignature-based IDSBehavioral ML platforms
Encrypted trafficMeasures behaviorCannot inspectVaries
Zero-day attacksDetected as deviationRequires signatureVaries
Training requiredNoneContinuous signature updatesDays to weeks
Independent benchmark validationPublished, reproduciblen/aGenerally not published
GPU requiredNoNoTypically yes
Air-gapped capableYesVariesGenerally no
Engine size500–560 KBHundreds of MBGB-scale infrastructure
Forensic signature per alert28 bytesLimitedVaries
Deployment specifications

500 KB. Air-gapped. Two operating modes.

Precision mode
500 KB engine
~1 ms per flow
3.8% FPR at α=0.05
CPU only — no GPU
Detection mode
560 KB engine
~4 ms per flow
14/15 classes ≥ 93% recall
Genesis S2 rescue for mimicry attacks
Deployment
Customer hardware — data never leaves
Air-gapped — no internet required
<30 sec / mins calibration by mode
No attack labels for calibration or detection
Operating envelope

From Raspberry Pi to enterprise server.

The same 500–560 KB engine runs everywhere Python 3.9+ runs. Throughput scales with hardware. Latency is ~1 ms per flow in Precision mode and ~4 ms in Detection mode on workstation-class machines.

Edge
~1,200 flows/sec
Raspberry Pi 5

Industrial sensors. Remote sites. Air-gapped enclaves.
Workstation
~4,800 flows/sec
Apple M2 / M4

SOC analyst workstations. Mid-tier deployments. Development environments.
Server
~15,000 flows/sec
32-core enterprise server

Production traffic monitoring. Enterprise-scale deployments.
Get started

See it on your traffic.

Reproduce the v2.1 benchmarks on corrected public datasets. Request a technical briefing. Evaluate VERITY on your own traffic, on your own hardware, with your team operating the engine throughout.