engram

Inspiration

Medical data is incredibly sensitive, yet accessing it securely across healthcare systems remains difficult. Traditional authentication relies on passwords, tokens, or centralized identity systems that don't necessarily provide strong privacy guarantees.

We wanted to explore a different approach: what if your brainwaves could become a privacy-preserving authentication mechanism?

Engram was inspired by the intersection of neuroscience, biometric authentication, and cryptography, with the goal of enabling secure access to medical data without compromising the privacy of the underlying biometric information.

What it does

Engram explores EEG-based biometric authentication combined with privacy-preserving cryptography to enable secure access to sensitive medical information.

Our system uses EEG signals, particularly Steady-State Visual Evoked Potentials (SSVEPs), to generate subject-specific biometric representations. These representations can be used to authenticate individuals while exploring ways to protect biometric templates through Fully Homomorphic Encryption (FHE).

The long-term vision is an authentication layer where users can prove their identity through brainwave responses without exposing their raw neural data or sensitive biometric templates.

How we built it

We built an EEG biometric authentication pipeline using PyTorch and SSVEP datasets, experimenting with neural embedding models and ArcFace-inspired metric learning to learn subject-specific representations.

We evaluated cross-session authentication and held-out identity generalization using the eldBETA dataset, investigating how stable and discriminative EEG representations remain across different recording sessions.

Alongside the biometric pipeline, we explored Fully Homomorphic Encryption using TenSEAL and CKKS to investigate encrypted biometric template operations and privacy-preserving verification.

Challenges we ran into

EEG signals are inherently noisy, highly variable, and sensitive to recording conditions. Learning representations that remain consistent across sessions while distinguishing different individuals was one of our biggest challenges.

Another challenge was balancing biometric accuracy with privacy. Encrypted computation introduces additional computational complexity, making it difficult to directly apply conventional authentication pipelines.

We also had to navigate the limitations of existing EEG datasets, including differences in recording sessions, channel configurations, and subject variability.

Accomplishments that we're proud of

We successfully trained EEG biometric embedding models and evaluated them on identities excluded from training.

Our ArcFace-based authentication experiments achieved a ROC-AUC of approximately 0.986 on a held-out identity evaluation, demonstrating promising discrimination between genuine and impostor EEG comparisons.

We also developed a working experimental pipeline for encrypted biometric template operations using CKKS, laying the groundwork for combining neural authentication with privacy-preserving computation.

Most importantly, we moved beyond the idea stage and built an experimental foundation for EEG-based authentication.

What we learned

We learned that EEG biometrics are much more challenging than conventional biometric modalities, but also potentially much more interesting.

We explored how metric learning can improve subject-specific representations, how cross-session variability impacts authentication, and how cryptographic techniques can be incorporated into biometric systems.

We also learned that privacy cannot simply be added at the end of an authentication pipeline. It needs to be considered throughout the system's design.

What's next for engram

Our next steps are to improve cross-session robustness, explore more efficient encrypted verification, and investigate real-time EEG acquisition for practical authentication.

We want to move toward a system that integrates wearable EEG hardware, privacy-preserving biometric verification, and secure medical data access.

Ultimately, we envision Engram as a privacy-first authentication infrastructure for sensitive healthcare data, where neural signals enable secure access without turning the brain itself into another source of exposed personal information.

Built With

• artificial
• biometrics
• cryptography
• data
• eeg
• fhe
• healthcare
• intelligence
• medical
• neuroscience
• privacy