Learning Latent Neural Signatures from EEG Scalograms for Robust Biometric Identification

Problem Statement

While Electroencephalography (EEG) offers a promising avenue for secure biometric identification due to the difficulty of replicating neural signals, its practical application faces significant hurdles. The primary challenges include the low signal-to-noise ratio, high inter- and intra-subject variability, and sensitivity to environmental fluctuations. Furthermore, existing deep learning approaches often treat EEG channels as independent temporal sequences, failing to explicitly preserve the spatial organization of the scalp electrode layout, which contains discriminative biometric information. Additionally, reliance on single-segment predictions makes systems vulnerable to transient artifacts (e.g., eye blinks, muscle activity), leading to unreliable identification decisions.

Methodology

The proposed framework reframes EEG-based biometric identification as a high-dimensional computer vision problem through a multi-stage pipeline:

1. Signal Preprocessing and Transformation:

   • Raw 64-channel EEG signals from the PhysioNet Motor Imagery dataset (109 subjects) are bandpass filtered (1–45 Hz) and resampled to 256 Hz.
   • Continuous recordings are segmented into non-overlapping 4-second windows.
   • Each segment undergoes a Continuous Wavelet Transform (CWT) using a complex Morlet wavelet to generate time-frequency scalograms, capturing the non-stationary spectral dynamics of the signal across 32 scales.

2. Latent Feature Compression:

   • To address computational complexity and overfitting risks associated with high-dimensional scalograms, a Convolutional Autoencoder is employed.
   • The encoder compresses each high-dimensional scalogram into a dense 16×16 latent patch. The decoder reconstructs the original signal to ensure feature fidelity, confirming that essential spatiotemporal information is retained in the compressed representation.

3. Spatial Grid Construction (2D CLSM):

   • The 64 latent patches are spatially arranged onto an 11×11 grid according to the international 10–20 electrode placement system.
   • This creates a 2D Cortical Latent Scalogram Map (2D CLSM) with dimensions of 176×176. Unoccupied grid cells are zero-padded. This step preserves the relative spatial organization and neighborhood relationships of the EEG channels, enabling the model to learn local feature interactions between adjacent cortical regions.

4. Deep Classification:

   • A custom Residual Neural Network (ResNet-18) is trained on the 2D CLSM images to extract subject-specific patterns from resting-state EEG. The architecture utilizes skip connections to mitigate vanishing gradients and learn deep hierarchies of brain features.

5. Session-Level Majority Voting:

   • To mitigate the stochastic nature of EEG and transient noise, the system employs a session-level majority voting strategy.
   • Instead of relying on a single 4-second segment, the model generates predictions for all 31 segments within a recording session. The final identity is determined by the mode (most frequent prediction) of these segment-level outputs. This aggregates temporally consistent neural evidence, reinforcing the correct identity even if individual segments are misclassified.

Key Contributions

• 2D Cortical Latent Scalogram Map (CLSM): Introduction of a novel feature engineering pipeline that transforms 1D raw EEG signals into 2D images. This involves computing CWT scalograms, compressing them via an autoencoder, and arranging them on a grid to preserve cortical spatial topology.
• Custom Deep Residual Learning: Design and training of a ResNet-18 architecture specifically tailored to capture subject-specific distinctive features from the constructed 2D CLSM representations.
• Session-Level Majority Voting: Implementation of a temporal aggregation strategy that significantly improves robustness by combining predictions across multiple segments, effectively filtering out noise-induced errors in isolated segments.

Results

The system was evaluated on the PhysioNet Motor Imagery dataset comprising 109 subjects.

• Segment-Level Performance: The ResNet model achieved a baseline accuracy of 87.1% when classifying individual 4-second EEG segments. The gap between training and validation accuracy highlighted the inherent stochasticity of single segments.
• Session-Level Performance: By applying the majority voting strategy across the 31 segments per session, the identification accuracy improved significantly to 99.1%.
• Visualization: t-SNE visualizations of the feature embeddings demonstrated strong inter-subject separation and compact intra-subject clusters, confirming that the model learns stable, unique neural signatures.
• Comparison: The proposed method outperformed several state-of-the-art approaches (including MLPs, XGBoost, and standard CNNs) that utilized different datasets or active motor tasks, achieving the highest reported accuracy on the large-scale PhysioNet resting-state dataset.

Significance and Claims

The paper claims that the proposed framework successfully advances EEG-based biometrics toward scalable, accurate, and secure real-world identification systems. By reformulating neural signal classification as a computer vision task, the study demonstrates that:

• Resting-state EEG contains stable, discriminative neural signatures: While individual segments may be noisy, the aggregated temporal evidence reveals a unique "neural fingerprint" for each subject.
• Spatial preservation is critical: Converting 1D signals into 2D maps that respect scalp topology allows deep learning models to effectively capture spatial relationships often missed by 1D-CNNs.
• Robustness through aggregation: The session-level voting strategy effectively mitigates the impact of transient artifacts, making the system reliable despite the stochastic nature of biological signals.

The authors note that the work aligns with SDG-3 (Good Health and Well-being) by enabling secure access to digital health systems and supporting potential applications in personalized neurophysiological monitoring. However, they acknowledge limitations, including the need for validation in dynamic scenarios (e.g., during motor tasks) and the computational cost of the current pipeline for edge device deployment.