engram

Inspiration

The idea behind engram started with a question: what if your brain activity could become a biometric?

We use fingerprints, faces and voices to identify people, but I wanted to explore something less conventional: the electrical activity of the brain.

I became particularly interested in SSVEP (Steady-State Visually Evoked Potentials), where the brain produces measurable responses to visual stimuli flickering at specific frequencies. I wanted to investigate whether these responses contain enough individual-specific information to distinguish one person from another.

What it does

engram is an experimental EEG-based biometric authentication system.

It uses deep learning to transform EEG recordings into 128-dimensional identity embeddings. These embeddings can then be compared to determine whether two recordings belong to the same individual.

Unlike a conventional SSVEP classifier that identifies which frequency a person is responding to, engram focuses on learning individual identity representations.

The goal is to investigate whether brain responses can provide a new form of biometric authentication.

How we built it

I built the system using Python and PyTorch, adapting an ArcFace-inspired metric learning architecture for EEG signals.

The experiments use the eldBETA dataset, containing EEG recordings from 100 participants, with 64 channels sampled at 1000 Hz.

I experimented with different electrode configurations, EEG preprocessing pipelines, embedding architectures and training strategies.

A major part of the work involved designing identity-separated evaluation experiments and testing whether learned representations could generalize across recording sessions.

I also explored Fully Homomorphic Encryption (FHE) as a potential way to protect biometric representations during future verification workflows.

Challenges we ran into

EEG is inherently noisy and non-stationary. The same person's brain activity can vary considerably between recording sessions.

Getting a model to distinguish individuals without simply memorizing recording-specific patterns was one of the biggest challenges.

Another challenge was designing meaningful evaluation protocols. Randomly splitting EEG samples can produce misleading results, so I focused on separating identities and recording sessions to investigate generalization more carefully.

Balancing the number of EEG channels, model complexity and available computational resources was another part of the experimentation.

Accomplishments that we're proud of

• Built an end-to-end experimental EEG biometric authentication pipeline.
• Adapted ArcFace-inspired metric learning for EEG identity embeddings.
• Experimented with multiple electrode configurations, including 6, 8, 16 and 32 channels.
• Evaluated identity verification using held-out identities and recording sessions.
• Achieved approximately 95% TAR at a calibrated operating threshold in one held-out evaluation experiment.
• Explored FHE for privacy-preserving biometric applications.

Most importantly, I moved from an initial research question to a working experimental authentication system.

What we learned

I learned that EEG biometrics are considerably more challenging than ordinary classification problems.

A model can achieve impressive training accuracy while still struggling with recordings collected under different conditions.

Metric learning, careful data splitting, threshold calibration and session-independent evaluation are essential when investigating biometric systems.

I also learned that accuracy is only one part of building a biometric system. Privacy, reproducibility, robustness and resistance to spoofing need to be considered before practical deployment.

What's next for engram

The next major step is collecting my own EEG recordings to evaluate the system beyond publicly available datasets.

I want to investigate how identity embeddings behave across different days, electrode placements, hardware configurations and recording conditions.

I'm also exploring encrypted biometric representations using Fully Homomorphic Encryption, with the longer-term goal of enabling verification without exposing sensitive biometric data.

Ultimately, I want to investigate whether brain activity can become a practical additional layer of digital identity.

engram is still an ongoing independent research project. There are many unanswered questions, and I intend to keep exploring them.

Built With

• arcface
• biometrics
• deep-learning
• eeg
• fastapi
• fhe
• machine-learning
• neuroscience
• numpy
• python
• pytorch
• scipy
• signal-processing