New EEG Dataset Enables Brain-Based Biometric ID Without External Cues

A new approach to EEG-based biometric identification uses imagined speech as a password-like security measure, removing the need for external visual or auditory guidance during data collection.

The Innovation#

Traditional EEG imagined speech studies rely on cues—such as flashing text or spoken prompts—to guide subjects through the cognitive task. This research departs from that model by letting participants naturally imagine words from a predefined list without any external triggers. The cueless paradigm more closely mimics real-world conditions where a user would voluntarily authenticate themselves, making it more practical for applications like brain-computer interfaces (BCIs).

The Dataset and Methods#

The team assembled a dataset of over 4,350 trials recorded from 11 subjects across five separate sessions. They tested multiple classification approaches, ranging from traditional machine learning—Support Vector Machines (SVM) and XGBoost—to modern deep learning models like EEG Conformer and Shallow ConvNet. A rigorous session-based hold-out validation strategy prevented the classifier from accidentally learning patterns specific to individual recording sessions rather than genuine subject identity.

Results#

The best models achieved 97.93% accuracy at distinguishing between subjects based solely on their brain activity during imagined speech. This level of performance suggests that neural signatures associated with language production are sufficiently unique and stable to serve as a reliable biometric marker.

Why It Matters#

The cueless approach addresses a major gap in brain-biometric research by showing that high-performance identification is possible under more naturalistic conditions. If validated further, such methods could enable secure, non-invasive authentication in BCI systems and other applications where brain signals encode personal identity information.