Deterministic swarm verdicts that cut IDS false positives, with signed audit trails. If your SOC drowns in alerts, this shows a measured way to combine detectors into verdicts you can replay and audit — including the honest record of where the approach hit its limits.
Who it's for: SOC analysts fighting alert fatigue · detection engineers · researchers who need falsifiable IDS baselines · agent builders who need runtime guardrails (OWASP ASI).
git clone https://github.com/Gautham-AI-dev/ai-network-auditor
cd ai-network-auditor
pip install -r requirements.txt
kaggle datasets download -d chethuhn/network-intrusion-dataset \
-p sentinel/data --unzip # one-time (free Kaggle account)
python reproduce.py # exit 0 = ALL MATCH (PASS)- 4 heterogeneous detector agents (rules, anomaly ×2, supervised) on a locked 586k-flow CICIDS2017 subset (DDoS + PortScan + benign).
- Deterministic adjudication: same inputs → byte-identical verdicts.
- Signed hash-chained audit log for every verdict (replay-verified).
- Red-team flood trials + graceful-degradation + determinism ablations.
- Full honesty record: the false-positive-reduction claim was FALSIFIED three ways with proof — the auditability that survived is the contribution.
- Verdict never performs worse than the best lone detector (guardrail holds).
- 10× duplicate-alert flood: zero accuracy loss, sub-second.
- Bit-identical repeat runs (chain fingerprints match exactly).
- Novelty, targets, and failures:
docs/novelty-lock.md(frozen).
flowchart LR
NET[Internet / WAN] --> FW[Firewall]
FW --> CORE[Core switch<br/>span / mirror port]
CORE --> SEN[Sentinel sensor<br/>passive tap, no inline risk]
SEN --> DET[Detector swarm<br/>rules + anomaly x2 + supervised]
DET --> VER[Deterministic verdict<br/>no AI in the verdict path]
VER --> CHAIN[(Signed audit chain<br/>replay-verifiable)]
VER --> SIEM[SIEM / SOAR<br/>only verdicts, not noise]
SIEM --> SOC[SOC analyst<br/>fewer, provable alarms]
RED[Red-team loop] -. attacks .-> DET
DRIFT[Drift monitor] -. watches .-> DET
The sensor is passive (a copy of traffic, never inline), so deploying it cannot break the network. Analysts stop drowning because only adjudicated verdicts — each with linked evidence — reach the SIEM, instead of every detector shouting separately.
flowchart TD
F[Network flows<br/>586k locked subset] --> D1[Rules agent]
F --> D2[Anomaly agent]
F --> D3[PCA agent]
F --> D4[Supervised agent]
D1 & D2 & D3 & D4 --> V[Verdict function<br/>fixed logic, train-locked]
V --> C{Meets bar?}
C -->|yes| OK[Verdict + evidence links]
C -->|no| RB[Auto-rollback<br/>to last stable config]
OK --> H[Hash chain append]
H --> AUD[Auditor replays<br/>any verdict offline]
intrusion detection, IDS, false positive reduction, alert correlation, SOC, SIEM, SOAR, threat hunting, CICIDS2017, AI security, AI attacks, AI defence, AI networks, agentic AI security, LLM security, prompt injection defense, agent oversight, OWASP agentic threats, audit trail, reproducible research, Python, machine learning security.
⭐️ If this helped your SOC or your research, star the repo — it keeps falsifiable security work visible.