HMOG Behavioral Biometrics Authentication System

A comprehensive authentication system using Hand Movement, Orientation, and Grasp (HMOG) behavioral biometrics with two different approaches.

Dataset

Download the HMOG dataset from: https://github.com/hmog-dataset/hmog

Place the dataset in hmog_dataset/public_dataset/ directory.

Setup

# Create virtual environment
python -m venv behav_env
source behav_env/bin/activate  # On Windows: behav_env\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Run the system
python main.py

Module Descriptions

1. data_processor.py - Handles data extraction and loading from user ZIP files
2. feature_extractor.py - Extracts HMOG and keystroke behavioral features
3. feature_processor.py - Performs feature engineering, selection, and preprocessing
4. authenticator.py - Traditional One-Class SVM authentication approach
5. global_feature_model.py - Global feature extractor and per-user classifiers
6. embedding_authenticator.py - Novel embedding-based authentication system
7. visualization.py - Generates results visualization and comparison plots
8. main.py - Orchestrates the entire authentication pipeline

How the Embedding-Based Approach Works

Core Concept

Instead of training separate models for each user on raw features, we:

1. Train a Global Model: Learn general behavioral patterns from all users
2. Extract Embeddings: Use the global model as a frozen feature extractor
3. Train Lightweight Classifiers: Each user gets a simple binary classifier

Benefits

• Better Generalization: Global model learns robust behavioral representations
• Less Data Required: Per-user classifiers need fewer samples
• Transfer Learning: Knowledge from all users helps individual authentication
• Computational Efficiency: Lightweight classifiers are fast to train/deploy

Pipeline

Raw Features → Global Model → Embeddings → User Classifier → Decision
    (X)            (frozen)       (128-D)      (SVM/RF/MLP)    (genuine/impostor)

Output

The system generates:

1. Console Output: Detailed progress and metrics
2. Visualizations:
   • traditional_results.png: Traditional approach performance
   • embedding_results.png: Embedding-based performance
   • comparison_results.png: Side-by-side comparison
   • per_user_comparison.png: Individual user comparison
3. Cached Files:
   • processed_data.pkl: Processed dataset
   • global_model.pkl: Trained global feature extractor

Performance Metrics

• EER (Equal Error Rate): Primary authentication metric
• Genuine Scores: Authentication scores for legitimate users
• Impostor Scores: Authentication scores for impostors
• Per-User EER: Individual user performance

References

If you use this system, consider citing relevant HMOG research papers on behavioral biometrics and authentication.