NeuroShield: A Device-Agnostic Foundation Model for EEG Authentication

Technical Summary: NeuroShield: A Device-Agnostic Foundation Model for EEG Authentication

Problem Statement
Current EEG authentication systems face a fundamental limitation: they are tightly coupled to specific acquisition settings, including fixed headset hardware, channel layouts, and signal durations. Existing models typically assume a predefined channel set, stable channel order, and fixed temporal windows. Consequently, variations in hardware (e.g., different electrode counts or placements) or recording protocols (e.g., varying signal lengths) render these models inapplicable without retraining. This fragmentation forces researchers to treat each new dataset or headset as a separate modeling problem, hindering knowledge transfer, limiting multi-dataset learning, and reducing model reusability. The field currently relies on dataset-specific pipelines rather than reusable encoders capable of accumulating knowledge across heterogeneous recording conditions.

Methodology: NeuroShield Architecture
To address these constraints, the authors introduce NeuroShield, a foundation model designed to learn identity-discriminative embeddings from variable-channel and variable-length EEG recordings. The core innovation is a dual-stage transformer architecture that decouples temporal modeling from spatial aggregation, allowing for flexible inputs.

1. Input Representation: EEG segments are represented as $S \in \mathbb{R}^{C \times T}$, where $C$ is the number of available channels and $T$ is the number of temporal samples. The model supports variations in both $C$ and $T$.
2. Temporal Patching & Encoding:
   • Each channel's waveform is divided into fixed-length temporal patches.
   • A Patch Encoder maps these raw signal fragments into a shared latent space.
   • A Temporal Transformer processes the sequence of patch tokens for each channel independently. It utilizes ALiBi (Attention with Linear Biases) to encode relative positional information, enabling the model to handle variable sequence lengths and extrapolate to longer durations than seen during training. A temporal classification token (CLS) is extracted for each channel to summarize its temporal features.
3. Spatial Aggregation:
   • The channel-level embeddings are aggregated using a Channel Transformer.
   • Crucially, this stage incorporates geometry-aware positional embeddings derived from the 3D coordinates of the electrodes. This allows the model to understand spatial relationships between electrodes without assuming a fixed channel order or layout.
   • A channel CLS token summarizes the final segment representation.
4. Training Objective: The model is trained using Supervised Contrastive Loss (SupConLoss). This objective encourages embeddings from the same subject to cluster together while pushing embeddings from different subjects apart, optimizing the metric space for authentication.

Experimental Protocol

• Pretraining: NeuroShield was pretrained on three large public datasets (TUH-EEG, DVS-rsEEG, PREDICT) comprising 15,762 subjects and 28,116 sessions. These datasets span multiple channel layouts (17–64 channels) and acquisition systems.
• Evaluation: The model was evaluated on two unseen downstream datasets: PEERS (345 subjects, 89 channels, medical-grade) and FRC-EEG (30 subjects, 14 channels, consumer-grade Emotiv Epoc X).
• Baselines: Comparisons were made against training the architecture from scratch (NS-Scratch) and a reproduced state-of-the-art baseline (Fallahi et al., 2026), which uses a ResNet1D encoder with fixed inputs.
• Settings: Evaluations included zero-shot transfer (no fine-tuning), full fine-tuning (NS-FullFT), partial fine-tuning (NS-LastFT), and parameter-efficient fine-tuning (NS-PEFT).

Key Results

• Performance Improvement: After fine-tuning, NeuroShield consistently outperformed the state-of-the-art baseline. On the PEERS dataset, it reduced the Equal Error Rate (EER) from 10.25% to 7.71% (a 2.54 percentage-point gain). On FRC-EEG, it reduced EER from 18.18% to 14.75%.
• Zero-Shot Capability: The pretrained model demonstrated transferable identity information in zero-shot settings, achieving EERs significantly below chance levels (e.g., 16.79% on PEERS) without any target-dataset updates, though fine-tuning was required for competitive performance.
• Temporal Flexibility: The model successfully generalized to input durations longer than those seen during training (up to 4.0 seconds vs. a 2.0-second training limit). Performance improved as more temporal context became available, with the 4-second PEERS result (5.47% EER) surpassing prior state-of-the-art results.
• Spatial Flexibility & Robustness:
  • The model operates effectively across different channel layouts and does not rely on fixed channel ordering; performance remained stable under channel permutations.
  • It demonstrated robustness to missing channels, tolerating up to 10–30% channel loss at verification time with only gradual performance degradation.
  • It generalized to entirely unseen electrode positions (channels not present in training data) without significant statistical difference in performance compared to seen channels.
• Data Composition: Analyses indicated that multi-session diversity in training data is more valuable than simply increasing the number of single-session samples for learning stable subject representations.

Significance and Claims
The paper claims that NeuroShield establishes a new paradigm for EEG authentication by shifting from isolated, dataset-specific models to a reusable, adaptable foundation model. Its significance lies in:

1. Breaking Hardware Dependencies: It enables authentication across heterogeneous recording settings (different headsets, channel counts, and layouts) without requiring separate model development for each setup.
2. Knowledge Accumulation: It demonstrates that identity-discriminative features can be learned across diverse datasets and transferred to unseen devices, supporting the accumulation of knowledge across studies.
3. Practical Deployment: By tolerating variable signal durations and missing channels, NeuroShield addresses real-world deployment challenges where recording conditions are rarely static or perfect.
4. Open Science: The authors release the model and evaluation pipeline as open source to facilitate reproducibility and community adoption.

The authors conclude that while NeuroShield reduces error rates relative to the state of the art, further work is needed to reach the maturity of established biometric modalities like face recognition. They suggest that future improvements may come from larger pretraining corpora, longer continuous segments, and more diverse consumer-grade device evaluations.