TonnieD/Mpesa-Fraud-Detection-System

Using Machine Learning to identify and prevent fraudulent transactions # M-Pesa Fraud Detection System A real-time transaction fraud detection and prevention system built for the Kenyan mobile money ecosystem. The system intercepts M-Pesa transactions before settlement and returns an `ALLOW`, `CHALLENGE`, or `BLOCK` decision within the latency window of a transaction initiation. --- ## Problem Statement M-Pesa processes **37.15 billion transactions annually**, with a total value of **KSh 38.29 trillion**. Conservative estimates place annual fraud losses at **KSh 2.3 billion (~$17.6M USD)** (approximately KSh 6.3 million stolen every single day). Common attack vectors include SIM swaps, social engineering, fraudulent merchant transactions, number masking, and fake balance SMS traps. --- ## Solution Rather than logging fraud after the fact, this system sits **between transaction initiation and transaction completion**, scoring each transaction in real time and returning one of three decisions: | Decision | Condition | |---|---| | `ALLOW` | Low fraud probability: transaction proceeds normally | | `CHALLENGE` | Elevated risk: OTP or user verification triggered | | `BLOCK` | High fraud probability or deterministic fraud signal: transaction halted before settlement | --- ## Project Structure ```text Mpesa-Fraud-Detection-System/ ├── Data/ │ ├── mpesa_synthetic.csv: Raw dataset (120K synthetic M-Pesa transactions) │ ├── Feature_engineered.csv: Engineered dataset │ ├── training.csv: Training split (109,915 rows) │ └── evaluation.csv: Unseen evaluation set (10,000 rows) ├── inference/ │ ├── models/ │ │ ├── best_model.pkl: Trained XGBoost model (GridSearchCV tuned) │ │ └── encoder.pkl: Fitted ColumnTransformer encoder │ ├── utils/ │ │ ├── feature_engineering.py: Derives drain_rate, account_emptied, cyclic encoding │ │ └── preprocessing.py: Drops columns, maps device_type, applies encoder │ ├── main.py: FastAPI application │ ├── requirements.txt: Production dependencies │ ├── .env: Local environment v …