Behavioral Biometric Authentication
LiveOverview
This project implements a behavioral biometric authentication system that identifies users from walking patterns using smartphone accelerometer and gyroscope data. Instead of passwords or fingerprints, the system explores passive authentication through gait recognition.
The project uses multi-sensor fusion, feature extraction, MLP classification, cross-day validation, and systematic optimization to reduce Equal Error Rate from baseline to 3.43%.
Architecture
Challenges & Solutions
Gait patterns change between collection days and walking conditions, making cross-day generalization hard.
Evaluated cross-day performance explicitly, using separate training and testing days to test real-world authentication robustness.
Single-sensor models underperform for gait-based authentication.
Combined accelerometer and gyroscope signals with 112 extracted features per window, achieving a best EER of 3.43%.
Results
Best EER
3.43%
Equal Error Rate (lower is better)
Accuracy
92.68%
MLP cross-day classification
Features
112
Time and frequency domain per segment
Dataset
10 users
2 collection days, 5s windows at 30Hz
Features
- Accelerometer and gyroscope multi-sensor fusion
- 112 time-domain and frequency-domain features per 5-second window
- MLP-based classification with architecture comparison
- Cross-day validation for real-world robustness testing
- Gait vs non-gait segment analysis
- Fisher Score feature selection - top-65 features yield best EER
- Equal Error Rate optimization workflow
