This project implements a comprehensive multi-user authentication system using neural networks for analyzing accelerometer-based behavioral biometrics. The system processes acceleration data from 10 different users and employs both pre-optimized and post-optimized neural network architectures to achieve robust user authentication.

• Develop a robust multi-user authentication system using accelerometer data
• Compare performance between pre-optimized and post-optimized neural network architectures
• Analyze behavioral biometric patterns across different data domains (time, frequency, and combined)
• Implement comprehensive performance evaluation metrics for authentication systems

PUSL3123-Coursework/ ├── README.md # This file ├── PUSL3123-Coursework-2024.pdf # Project specification document ├── (01)-Data-Analysis/ # Initial data analysis scripts │ ├── load_data.m # Data loading utility │ ├── analyze_inter_variance.m # Inter-user variance analysis │ └── analyze_intra_variance.m # Intra-user variance analysis ├── (02)-Pre-optimized-Neural-Network/ # Pre-optimization implementation │ ├── main.m # Main execution script │ ├── 1_data_analysis/ # Data analysis modules │ ├── 2_neural_network/ # Neural network implementation │ │ ├── prepare_data/ # Data preparation utilities │ │ ├── train/ # Training modules │ │ ├── evaluate/ # Evaluation utilities │ │ └── results/ # Output results │ └── CW-Data/ # Dataset (60 .mat files) └── (03)-Post-optimized-Neural-Network/ # Post-optimization implementation ├── main.m # Main execution script ├── 1_data_analysis/ # Enhanced data analysis ├── 2_neural_network/ # Optimized neural network │ ├── prepare_data/ # Enhanced data preparation │ ├── train/ # Optimized training algorithms │ ├── evaluate/ # Advanced evaluation metrics │ └── results/ # Comparative results └── CW-Data/ # Dataset (60 .mat files) 

The dataset consists of accelerometer-based behavioral biometric data from 10 users (U01-U10). Each user has 6 different data files representing different feature extraction methods and collection scenarios:

• Time Domain (TimeD): Raw accelerometer signals in time domain
• Frequency Domain (FreqD): Frequency-transformed features
• Combined (TimeD_FreqD): Hybrid time-frequency features

• Same Day (FDay): Training and testing data collected on the same day
• Cross Day (MDay): Training and testing data collected on different days

U[XX]_Acc_[Domain]_[Scenario].mat 

• XX: User ID (01-10)
• Domain: TimeD, FreqD, or TimeD_FreqD
• Scenario: FDay or MDay

Total Dataset: 60 .mat files (10 users × 6 data variations each)

• Architecture: Standard feedforward neural network
• Hidden Layers: Fixed architecture
• Activation Functions: Default MATLAB configurations
• Training Algorithm: Standard backpropagation

• Dynamic Architecture: Adaptive hidden layer sizing using multiple heuristics
• Optimized Training: Enhanced training algorithms with early stopping
• Advanced Metrics: Comprehensive authentication performance evaluation
• Regularization: Improved generalization techniques

1. Dynamic Neuron Calculation:

   • Geometric pyramid rule
   • Input size-based rule
   • Sample size-based rule
   • Median-based final selection

2. Training Enhancements:

   • Monitoring and early stopping
   • Adaptive learning rates
   • Cross-validation

3. Performance Metrics:

   • False Acceptance Rate (FAR)
   • False Rejection Rate (FRR)
   • Equal Error Rate (EER)
   • Area Under Curve (AUC)
   • F1-Score

• MATLAB R2019b or later
• Neural Network Toolbox
• Statistics and Machine Learning Toolbox

1. Clone or download the project repository
2. Open MATLAB and navigate to the project directory
3. Ensure all required toolboxes are installed

cd '(02)-Pre-optimized-Neural-Network' main

cd '(03)-Post-optimized-Neural-Network' main

cd '(01)-Data-Analysis' analyze_inter_variance analyze_intra_variance

When running the main scripts, you'll be prompted to select:

1. Time domain - Analysis using temporal features
2. Frequency domain - Analysis using spectral features
3. Combined - Analysis using both time and frequency features

The system evaluates authentication performance using industry-standard metrics:

• Accuracy: Overall classification accuracy
• FAR (False Acceptance Rate): Rate of incorrectly accepting impostors
• FRR (False Rejection Rate): Rate of incorrectly rejecting genuine users
• EER (Equal Error Rate): Point where FAR equals FRR
• AUC (Area Under Curve): ROC curve area measurement
• F1-Score: Harmonic mean of precision and recall

• Confusion matrices for each user
• ROC curves
• Performance comparison charts
• Training progress monitoring

• Inter-variance Analysis: Examines differences between users
• Intra-variance Analysis: Examines consistency within users
• Statistical Profiling: Comprehensive feature statistics

• Adaptive Architecture: Dynamic hidden layer sizing
• Multi-domain Support: Time, frequency, and combined features
• Cross-validation: Robust performance estimation
• Early Stopping: Prevents overfitting

• Comprehensive Metrics: Multiple performance indicators
• User-specific Analysis: Individual user performance profiling
• Comparative Analysis: Pre vs. post-optimization comparison
• Visual Reports: Automated result visualization

• Improved Authentication Accuracy: Post-optimized networks typically achieve 85-95% accuracy
• Reduced Error Rates: Lower FAR and FRR compared to pre-optimized versions
• Better Generalization: Enhanced performance on cross-day scenarios
• Optimized Architecture: Automatically tuned network parameters

• Time Domain: Good for temporal patterns
• Frequency Domain: Effective for spectral characteristics
• Combined Domain: Best overall performance
• Same Day: Higher accuracy due to reduced environmental variations
• Cross Day: More challenging but realistic scenario

• Automated data file discovery
• Structured data organization
• Memory-efficient loading

• Dynamic architecture calculation
• Configurable parameters
• Binary classification setup

• Adaptive training algorithms
• Progress monitoring
• Early stopping mechanisms

• Comprehensive metric calculation
• Statistical analysis
• Performance visualization

• Modular Design: Easy to extend and modify
• Error Handling: Robust error management
• Logging: Comprehensive execution logging
• Reproducibility: Consistent results across runs