Fraud Detection MLOps System

Real-time credit card fraud detection pipeline built with PyTorch, XGBoost, Apache Kafka, MLflow, and FastAPI , deployed with Docker

---

Results

Model	AUC-ROC	Precision	Recall	F1
PyTorch NN	0.9745	0.6569	0.9054	0.7614
XGBoost	0.9859	0.9403	0.8514	0.8936

• Results on Kaggle Credit Card Fraud dataset
• Optimising for Recall to minimise missed fraud cases
• Pytorch deployed as primary model
• XGBoost used as a confirmation Layer

---

Architecture

                    REAL-TIME INFERENCE PIPELINE            
                                                  
  Transactions → Kafka Topic → Consumer → FastAPI 
                                    ↓             
                                  Model Inference
                                               
                        
─────────────────────────────────────────────────────────────────────────

                    TRAINING PIPELINE                    
                                                         
  Raw Data → Preprocessing → Model Training  → MLflow Experiment Tracking                           
                                                        ↓
                                                      Model Registry

---

Tech Stack

Layer	Technology
ML Framework	PyTorch 2.0, XGBoost
Model Serving	FastAPI + Uvicorn
Streaming	Apache Kafka
Experiment Tracking	MLflow
Containerization	Docker + Docker Compose
Data Processing	Scikit-learn, Pandas, NumPy

---

Dataset :

https://www.kaggle.com/datasets/mlg-ulb/creditcardfraud

• Get creditcard.csv
• Place it at: data/creditcard.csv

---

Quick Start

Option 1 — Docker run (recommended)

git clone https://github.com/Strifee/fraud-detection-mlops

cd fraud-detection-mlops

# Start all services (Kafka, MLflow, API)
docker-compose up -d

# Check API health
curl http://localhost:8000/health

Option 2 — Local development

pip install -r requirements.txt

python data/preprocess.py

python model/train.py --model both

mlflow ui --port 5000
# Open http://localhost:5000

uvicorn api.main:app --reload --port 8000
# Open http://localhost:8000/docs


python streaming/kafka_simulator.py --mode simulate

---

API Endpoints

Model Health Check

Single transaction prediction

Interactive API docs

http://localhost:8000/docs

---

Model Details

PyTorch Neural Network

• Architecture: Input(30) → 512 → 256 → 128 → 64 → 1
• BatchNorm + Dropout (0.3) at each layer
• BCEWithLogitsLoss with positive class weighting for imbalance
• Adam optimizer with ReduceLROnPlateau scheduler

XGBoost

• 300 estimators, max depth 6
• scale_pos_weight=10 for class imbalance handling
• Trained with early stopping on validation AUC

Class Imbalance Strategy

• Dataset: 0.17% fraud rate (highly imbalanced)
• PyTorch: weighted loss function
• XGBoost: scale_pos_weight parameter
• Preprocessing: controlled undersampling for training balance

---

MLflow Experiment Tracking

All runs are tracked including:

• Hyperparameters
• Training metrics per epoch (loss, AUC, F1, Precision, Recall)
• Test set final metrics
• Inference latency
• Model artifacts

mlflow ui
# Open http://localhost:5000 to compare runs