mmachelane/sa-fraud-shield

Transaction Fraud Detection for South African Banking # 🇿🇦 sa-fraud-shield ### Production-Grade SA Fraud Detection — GNN + Streaming + Federated Learning + POPIA-Compliant XAI --- ## 🎯 Project Overview **Business Problem:** South African banks lost **R2.7 billion to financial crime in 2024**, with digital banking fraud surging 86% to ~98,000 incidents. Every R1 lost costs SA banks **R4.52** in total when factoring in investigation, recovery, and customer churn. Yet no SA bank has deployed graph neural networks, federated learning, or transformer-based models — they're still running FICO Falcon (circa 2011) and vendor-supplied solutions. **Solution:** A **production-grade fraud detection system** that demonstrates what SA banks actually need but haven't shipped yet: - **AUC 0.9922** SIM swap detection model — SA's dominant fraud vector (60% of mobile breaches) - **Graph Neural Network** fraud ring detector on heterogeneous transaction graphs - **Real-time Kafka streaming** with sub-100ms scoring latency - **Federated learning** across 5 simulated SA banks with differential privacy (POPIA-compliant) - **LLM-generated explanations** in English and isiZulu — satisfying POPIA Section 71 **Why This Matters:** SA banks plan to invest **R30M+ each in AI in 2026**. This project demonstrates the exact stack they need — and that no Kaggle notebook comes close to replicating. --- ## 📊 Key Results | Metric | Value | Context | |--------|-------|---------| | **SIM swap AUC** | **0.9922** | Temporal cross-validation, 34 features | | **GNN AUC** | 0.6450 | GraphSAGE on heterogeneous fraud graph | | **API latency** | = 0.7"] }, "latency_ms": 4.2 } ``` ### Run Tests ```bash pytest tests/ # 140 passed in 44s ``` --- ## 📈 Build Phases | Phase | Component | Highlight | |-------|-----------|-----------| | 1–2 | Scaffold + shared library | Pydantic schemas, SA validators, constants | | 3 | Synthetic data generation | SA-locale Faker, PayShap IDs, load shedding injection | | 4 | SIM swap model | LightGBM, 34 features, tempo …