🛡️
OUROBOROS
Enterprise Security Platform
Autonomous AI Defense Lab for Next-Generation Payment Networks
A closed-loop security system enabling financial institutions and payment processors to simulate emerging GenAI adversarial vectors, discover static rule evasion blind spots, and continuously retrain explainable ML models prior to production release.
● Adversary Simulation Active
Continuous ML Retraining
Protocol: Closed-Loop Zero-Trust Defense · Real-Time Explainability
14,000
Synthesized Attack & Baseline Events
100.0%
Detection Recall (@ 1% FPR)
1.0%
Strict False-Positive Budget
4 Vectors
Phish · Vish · Camouflage · Injection
🕹️ One-Click Adversary Simulation Presets (Instant In-Memory Cache)
Select Attack Vector to Synthesize / Inspect
💡 Clicking below transfers this exact generated transaction into the Live Detection tab to evaluate how the ML model scores it.
🎯 Interactive Triage Presets
📝 Transaction Parameters
Channel
💡 See how proactive synthetic generation enabled us to close the $3,000 threshold gap through retraining.
⚡ CLOSED-LOOP RESULT: +85.0% Recall Gain on Adaptive Adversaries
By proactively generating the camouflage evasion attack, we exposed the static threshold vulnerability and retrained our classifier with velocity features prior to production deployment.
PHASE 1 · BASELINE DEFENSE
🛡️ Static Rule ($3k)
High precision (97.5%) on traditional large transactions, but totally blind to multi-step micro-structuring.
PHASE 2 · THE ADVERSARIAL GAP
⚠️ 85% Camouflage Missed
Adversarial LLM split $15,000 into six $2,500 payments, evading single-transaction threshold rules completely.
PHASE 3 · CLOSING THE LOOP
🔄 Synthetic Feedback Retrain
Synthesized evasion sequences fed back into model training with 24h velocity & inter-transaction delta features.
PHASE 4 · HARDENED DEFENSE
🚀 100.0% Detection Recall
Production XGBoost classifier catches micro-structuring at a strict 1.0% FPR budget without human intervention.
📈 Overall Model Performance (@ 1.0% FPR Budget)
| Precision | 97.6% |
| Recall | 100.0% |
| F1-Score | 0.988 |
| AUC-PR | 0.9996 |
🎯 Test Set Confusion Matrix (4,200 Events)
| True Positives (Fraud Caught) | 1,200 |
| True Negatives (Legit Cleared) | 2,970 |
| False Positives (Legit Blocked) | 30 (1.0%) |
| False Negatives (Fraud Missed) | 0 |
System Framing
Ouroboros is an autonomous AI Defense Lab for payment networks that simulates generative adversarial fraud vectors, discovers evasion gaps in existing static rules, and retrains production ML classifiers to close those gaps before attacks reach production.
Attack Vectors Modeled
- Attack 1 (Phishing): Mass-personalized spear phishing generating context-aware SMS/email lures.
- Attack 2 (Vishing): Voice-cloned social engineering inducing authorized push payment (APP) transfers.
- Attack 3 (Camouflage): Adversarial temporal structuring/smurfing breaking large fraud transfers into sub-$3,000 bursts.
- Attack 4 (Agentic Injection): Prompt injection embedded in product listings hijacking autonomous purchasing agents.
Base Datasets & Real-World Feasibility
- Sparkov Fraud Dataset (CC0 Public Domain): Contributes cardholder transaction distributions, merchant categories, and diurnal spending patterns.
- PaySim Mobile Money Dataset (CC BY 4.0): Contributes mobile wire transfer velocity, balance dynamics, and multi-step cash-out topologies.
Architecture & Technical Stack
- Adversarial Generation: Google Gemini Pro / Gemma via Google AI Studio API for synthetic payload synthesis.
- Detection & Explainability: XGBoost primary classifier, Isolation Forest anomaly detector, and TreeSHAP feature attributions.
- Voice Channel Safety: Uses standardized synthetic voice personas for caller scripts, with zero real-world biometric or voice data collected.
Headline Result
- +85.0% Detection Recall Improvement: Proactive synthetic generation exposed an 85.0% evasion gap on camouflaged structuring attacks against static $3,000 threshold rules, retraining the model to achieve 100.0% recall at a strict 1.0% False Positive Rate (FPR) budget.
Project Resources & Links
- Architecture Documentation:
docs/architecture.md - Problem Framing:
docs/problem_framing.md - Known Limitations:
docs/known_limitations.md - Voice Channel Ethics Note:
docs/voice_channel_note.md
Base datasets: Sparkov Fraud Simulation (CC0 Public Domain) · PaySim Mobile Money (CC BY 4.0).
Built for the Mastercard Innovation Challenge 2026.