TL;DR
This paper presents a novel two-stage EEG-based authentication system that uses deep learning (Inception or VGG-like networks) or Principal Component Analysis (PCA) for feature extraction, followed by a Support Vector Machine (SVM) for classification. The system achieves robust authentication with high accuracy and low false acceptance rates, demonstrating the potential of EEG as a biometric for enhanced security.
What Problem Does This Paper Address?
In an era of increasing digital data, robust and secure authentication methods are paramount to prevent unauthorized access. While traditional biometrics like fingerprints and iris scans are common, they can be compromised or are not always practical. Electroencephalography (EEG), which captures brainwave activity, offers a unique and difficult-to-replicate biometric signature. However, existing EEG-based authentication systems often face challenges with computational complexity, lengthy evaluation times, and achieving high accuracy with single-trial data. This paper aims to address these limitations by developing a more efficient and accurate single-trial EEG authentication system.
Methods
The proposed system employs a two-stage classification approach.
Stage 1: Feature Extraction (Dimensionality Reduction) The raw EEG signals are first processed using Empirical Mode Decomposition (EMD) to extract Intrinsic Mode Functions (IMFs), which represent different frequency bands. From these IMFs, various features are computed, including Shannon entropy, log energy, approximate entropy, sample entropy, and power spectral density (PSD) for specific frequency bands (mu and beta rhythms). This results in an initial feature set of 18 features per channel.
To reduce the dimensionality of these features, three methods are compared:
- Principal Component Analysis (PCA): A classical dimensionality reduction technique that identifies uncorrelated components capturing the most variance.
- Inception-like Deep Neural Network (NN): A convolutional neural network architecture inspired by GoogleNet's Inception modules, designed to capture complex dependencies within features and between channels.
- VGG-like Deep Neural Network (NN): A convolutional neural network architecture inspired by VGGNet, characterized by its depth and use of small convolutional kernels.
These deep learning models are trained using a transfer learning approach, where they are initially trained on a larger dataset and then adapted for feature extraction in this specific authentication task. The output of this stage is a reduced feature set (typically 2 features per channel, concatenated to form a 36x1 vector).
Stage 2: Classification A Support Vector Machine (SVM) with a radial basis function (RBF) kernel is used as the binary classifier. The SVM is trained to distinguish between genuine users and imposters based on the features extracted in Stage 1. The system is designed for single-trial authentication, meaning it aims to authenticate a user based on a single EEG recording.
The researchers also explored the impact of using different numbers of EEG channels (8, 16, and 64) on performance.
Datasets
The study utilizes the publicly available "EEG Motor Movement/Imagery Dataset" from PhysioNet. This dataset contains EEG recordings from 105 subjects who performed cued motor-movement and imagery tasks. Each subject underwent multiple sessions and trials, with the data preprocessed to create approximately 105 trials per subject for each task. This dataset serves as both the training and testing ground for the proposed authentication system.
Key Insights and Ideas
- Single-trial EEG authentication is feasible and robust: The study demonstrates that even with a single EEG recording, a high degree of authentication accuracy can be achieved, making it practical for real-time applications.
- Deep learning architectures offer competitive feature extraction: Both Inception-like and VGG-like neural networks, when used for feature extraction, show promising results, highlighting their ability to learn complex patterns in EEG data.
- PCA remains a strong contender: Despite the advancements in deep learning, PCA-SVM still delivers competitive and, in some cases, superior results, especially with a higher number of channels, suggesting its efficiency and effectiveness for this task.
- More channels generally improve performance, but with diminishing returns for some models: Increasing the number of EEG channels generally leads to better accuracy and lower FAR, particularly for PCA and the Inception-like NN. However, the VGG-like NN's performance is less sensitive to the number of channels.
- The trade-off between complexity and performance: While deep learning models offer sophisticated feature extraction, they come with significantly longer training times compared to PCA. This is a crucial consideration for practical deployment.
Key Results
The study evaluated three systems: PCA-SVM, Inception-like NN-SVM, and VGG-like NN-SVM, across 8, 16, and 64 channels.
- Highest Accuracy: The PCA-SVM system achieved the highest accuracy, reaching 95.64% with 64 channels.
- Lowest False Acceptance Rate (FAR): The PCA-SVM system also demonstrated the lowest FAR, with 1.26% for 64 channels. The Inception-like NN-SVM also achieved a low FAR of 1.27% with 64 channels.
- Performance across channel counts:
- 8 Channels: PCA-SVM achieved 92.09% accuracy and 2.19% FAR. VGG-like NN-SVM achieved 88.29% accuracy and 2.55% FAR. Inception-like NN-SVM achieved 87.28% accuracy and 4.08% FAR.
- 16 Channels: PCA-SVM achieved 92.36% accuracy and 2.17% FAR. VGG-like NN-SVM achieved 87.47% accuracy and 3.33% FAR. Inception-like NN-SVM achieved 85.31% accuracy and 3.53% FAR.
- 64 Channels: PCA-SVM achieved 95.64% accuracy and 1.26% FAR. VGG-like NN-SVM achieved 90.68% accuracy and 2.89% FAR. Inception-like NN-SVM achieved 93.40% accuracy and 1.27% FAR.
- Statistical Significance: Statistical tests indicated that the Inception-like NN-SVM performed significantly worse than VGG-like NN-SVM and PCA-SVM for 8 and 16 channels in terms of FAR. For 64 channels, VGG-like NN-SVM was significantly outperformed by both PCA-SVM and Inception-like NN-SVM.
- Training Time: PCA-SVM required approximately 30 minutes for training, while VGG-like NN-SVM took about 1 hour, and Inception-like NN-SVM took around 3 hours on a standard laptop.
Limitations and Future Work
The authors acknowledge several limitations and suggest future research directions:
- Dataset Comparability: For a fair comparison with existing literature, it would be beneficial to test published methods on the dataset used in this study. The authors have made their code publicly available to facilitate this.
- Electrode Application Time: The application of EEG electrodes can be time-consuming, posing a hurdle for widespread adoption. Future work could investigate the use of dry electrode EEG recordings to streamline this process.
- User Training: Adding a new user to the system requires training a new authentication model for that individual.
- Robustness Testing: The system's robustness needs to be tested using data collected on different days and from various recording sessions to ensure consistent performance under diverse conditions.
- Real-world Deployment: While the PCA-SVM is recommended for its balance of accuracy, speed, and simplicity, further research into optimizing deep learning models for faster training and inference would be valuable for real-world applications.
Summary updated