AI/ML · Biometrics · Signal Processing
Acceleration-Based User Authentication
Completed
Behavioral biometric authentication from mobile accelerometer signals using a neural network — ~96.4% accuracy.
Overview
A complete behavioral-biometrics pipeline that verifies a user's identity from how they move. Raw accelerometer signals are segmented into windows, reduced to 33 time-domain features, normalized, and fed to a feed-forward neural network trained with cross-entropy and scaled conjugate gradient. The system is evaluated with confusion matrices, per-user FAR/FRR and EER, visualized with PCA, and tuned via a grid search over architectures — all reproducible end to end.
Key features
Feature extraction
33 time-domain features from windowed signals.
Normalization pipeline
Standardized feature scaling.
Neural network
Feed-forward classifier, cross-entropy, trainscg, 70/15/15 split.
Evaluation suite
Confusion matrix, per-user FAR/FRR & EER curves.
PCA visualization
2D scatter of feature separability.
Architecture optimization
Grid search over NN architectures (selected [30]).
Tech stack
Language
- MATLAB
ML
- Feed-forward NN (pattern recognition)
- trainscg
- Cross-entropy
Techniques
- Time-domain feature engineering
- Normalization
- PCA
- Grid search
Domain
- Behavioral biometrics
- Sensor/signal processing
Architecture
End-to-end pipeline: windowed segmentation → 33-D feature vector → normalization → NN classifier → biometric evaluation (EER/FAR/FRR). Architecture chosen by grid search balancing accuracy, variance, and speed.
Highlights
- ~96.4% authentication accuracy
- 33 engineered time-domain features
- Per-user EER & FAR/FRR evaluation
- Architecture [30] selected via grid search
- Fully reproducible (raw data → model → figures)