EEG-Based Biometric Authentication using Deep Learning

Overview

This project presents a EEG Biometric Authentication System that utilizes Digital Signal Processing (DSP) and Deep Learning (CNN) for secure user authentication. The system processes Electroencephalography (EEG) signals, extracts discriminative time-frequency features using the Continuous Wavelet Transform (CWT) and generate Scalogram images, and classifies registered users using a Convolutional Neural Network (CNN). It also supports unknown user rejection through a majority-voting authentication mechanism.

---

Features

• Multi-user EEG biometric authentication
• EEG preprocessing using DSP techniques
• Bandpass filtering (8–30 Hz Butterworth IIR Filter)
• Continuous Wavelet Transform (CWT)
• Scalogram-based feature extraction
• Patch-based image generation (avoids information loss)
• Multi-channel CNN classification
• Unknown user rejection
• GUI-based authentication system
• Performance evaluation using standard metrics

---

Project Workflow

EEG Signal
     │
     ▼
Epoch Segmentation
     │
     ▼
Normalization
     │
     ▼
Bandpass Filter (8–30 Hz)
     │
     ▼
Wavelet Transform (CWT)
     │
     ▼
Scalogram Generation
     │
     ▼
Patch Extraction
     │
     ▼
Multi-Channel Image Stacking
     │
     ▼
CNN Classification
     │
     ▼
Majority Voting
     │
     ▼
Authentication
(Authorized / Unauthorized)

---

Digital Signal Processing Pipeline

Preprocessing

• EEG Epoch Segmentation
• Z-score Normalization
• Butterworth Bandpass Filtering (8–30 Hz)

Feature Extraction

• Fast Fourier Transform (FFT)
• Spectrogram (STFT)
• Continuous Wavelet Transform (CWT)
• Scalogram Generation

After comparative analysis, Wavelet Transform with Scalogram representation was selected because it preserves both time and frequency information effectively.

Deep Learning Architecture

The proposed CNN consists of:

• Input Layer
• Convolution Layer (3×3)
• ReLU Activation
• Max Pooling Layer
• Convolution Layer
• Max Pooling Layer
• Flatten Layer
• Fully Connected Layer
• Dropout Layer
• Softmax Output Layer

---

Authentication Strategy

The trained CNN predicts the class of each extracted EEG patch.

The final authentication decision is obtained using majority voting:

• If the majority of patches belong to a registered user:
  • Authorized
• Otherwise:
  • Unauthorized

---

Dataset

Dataset: EEG Motor Movement/Imagery Dataset

Dataset Format:

• EDF (European Data Format)

Dataset Information:

• 64 EEG Channels
• Sampling Frequency: 160 Hz
• Multiple Subjects
• Multiple Recording Sessions

---

Installation

Clone the repository:

git clone https://github.com/Prince-yadav2/EEG-Biometric-Authentication.git

Go to the project directory:

cd EEG-Biometric-Authentication

Install the required libraries:

pip install -r requirements.txt

---

Running the Project

Run the GUI:

python GUI_Multi_User_Authentication.py

The GUI allows you to:

• Select an EEG (.edf) file
• Authenticate the user
• Display the authentication result

---

Performance Metrics

The proposed system is evaluated using:

• Accuracy
• Precision
• Recall (Sensitivity)
• Specificity
• F1-Score
• Matthews Correlation Coefficient (MCC)
• AUC-ROC
• False Acceptance Rate (FAR)
• False Rejection Rate (FRR)
• Confusion Matrix

Metric	Value
Accuracy	86.88%
Precision	79.74%
Recall	76.40%
Specificity	95.97%
F1 Score	77.57%
MCC	76.18 %
AUC-ROC	97.00 %
FAR	04.03 %
FRR	23.59 %

---

Results

Filter Comparison

---

User-Level Confusion Matrix

---

GUI

---

Future Work

• Improve unknown-user rejection
• Optimize CNN architecture
• Real-time EEG acquisition
• Hardware implementation using embedded systems
• Deploy the authentication system as a standalone desktop application

---

Author

Prince Yadav

Department of Electrical Engineering

---

Acknowledgements

• PhysioNet EEG Motor Movement/Imagery Dataset
• MNE-Python
• TensorFlow / Keras
• Scikit-learn
• PyWavelets

---

License

This project is intended for educational and research purposes.