Fraud Detection MLOps on GKE 🚀

real-time fraud detection system for online payment transactions.
This project demonstrates the full lifecycle of an ML system: from data preparation and model training to containerization, deployment, monitoring, and security.

---

📂 Repository Structure

• data/ → Versioned datasets (v0, v1, poisoned variants) tracked with DVC
• src/ → Core FastAPI service and model code
• feature_repo/ → Feast feature store setup
• .github/workflows/ → CI/CD pipelines with GitHub Actions + CML reporting
• deployment.yaml, service.yaml, hpa.yaml → Kubernetes manifests for GKE deployment and autoscaling
• locustfile.py → Load testing scripts
• Dockerfile → Containerization of FastAPI + model
• requirements-*.txt → Environment dependencies for API, training, CI/CD

---

📊 Dataset

• Source: European cardholder transactions (transactions.csv)
• Schema: 31 numerical columns (Time, V1–V28, Amount, Class)
• Class: 1 = fraud, 0 = non-fraud
• Preparation:
  • Split into data/v0/transactions_2022.csv and data/v1/transactions_2023.csv
  • Poisoned datasets created with flipped labels (2%, 8%, 20%)

---

🛠 Features

1. CI/CD & Containerization

• FastAPI /predict endpoint (returns fraud probability + prediction)
• Dockerized service pushed to Google Artifact Registry
• GitHub Actions workflow with CML reporting

2. Deployment & Scaling

• GKE deployment with LoadBalancer service
• HorizontalPodAutoscaler (HPA) for CPU-based scaling
• Load testing with Locust

3. MLSecurityOps

• Data poisoning attack simulation
• DVC for dataset versioning
• MLflow experiment tracking with poisoning-level parameter

4. Explainability & Fairness

• SHAP beeswarm plots for feature importance
• Fairlearn audits for demographic parity difference
• Concept drift detection between v0 and v1 datasets

---

📈 Monitoring & Observability

• OpenTelemetry instrumentation for API latency and model inference time
• MLflow logging of metrics, parameters, and artifacts
• Drift comparison plots for v0 vs v1 performance

---