Mutumades/behaviourial-biometrics

AI-driven behavioral biometric fraud detection system for mobile banking in Kenya — detects SIM-swap, RAT/device takeover, social engineering, bot-like attacks, and high-value mule fraud using touch, typing, navigation, and context patterns. 🧠 Behavioral Biometric Fraud Detection — Kenya Post-Login AI Security for Mobile Banking 🌍 Overview This project builds an AI-powered behavioral biometric security system to detect fraudulent sessions after login in mobile banking apps. Unlike traditional security (PINs, OTPs, facial/ fingerprint login), this system monitors how users behave during the session — their: 🖐️ Touch dynamics ⌨️ Typing patterns 🧭 Navigation flow 📍 Context (location, time, SIM, device, network) It can catch attackers who bypass login using Artificial Intelligence deepfakes, stolen OTPs, or remote access tools. ⚠️ Fraud Types Detected The system simulates and detects 5 major fraud categories common in Kenya and Africa: Fraud Type Description SIM-Swap Fraud Criminal hijacks the victim’s phone number to intercept OTPs Remote Access Fraud (RAT) Fraudster remotely controls the victim’s phone and drains funds Social Engineering Scams Victims are tricked into sending money to fraudsters Bot-like Behavior Automated scripts mimic human behavior to steal funds fast High-Value Mule Transfers Sudden large transfers far beyond a user’s normal behavior 🧪 Dataset Synthetic behavioral biometrics dataset of 700 users ~25–45 sessions each (heavy-tailed distribution) ~5% labeled fraud Includes touch, typing, navigation, and context data Realistic fraud behaviors injected (SIM/device swaps, remote flows, bots, outlier amounts) ⚡ This approach is ethical and realistic — real behavioral banking data is private and unavailable, so synthetic data is required for safe experimentation. ⚙️ Tech Stack Python, pandas, scikit-learn, xgboost matplotlib and seaborn for analysis & visualizations Streamlit + pyngrok for real-time fraud detection dashboard Google Colab for development 📊 Results Model Accuracy Precision Recall F1 ROC-AUC Logistic Regression 92.6% 99.3% 92.9% 96.0% 0.941 Random Forest 97.9% 97.9% 99.9% 98.9% 0.957 XGBoost 98.6% 98.7% 99.8% 99.2% 0.964 ✅ XGBoost was selected as the fin …