A personal research project exploring whether biometric authentication is possible without storing any biometric data — not even a derived template.

ZKHD combines three ideas:

SimHash encoding — ArcFace face embeddings (512-dim floats) are projected into 32,000-bit binary hypervectors via Sign Random Projections. The encoding preserves angular distance: same-person pairs land close in Hamming space, different-person pairs land near 0.5.

Fuzzy extraction — a cryptographic primitive that derives a stable key from a noisy biometric using stored helper data. Currently mocked with an oracle; the real construction requires an error-correcting code (SC-LDPC) that this project has not yet implemented.

Zero-knowledge proof — the user should prove "my biometric is close enough" without revealing anything. Currently mocked with an Ed25519 signature.

The representation layer (ArcFace + SimHash) is implemented and empirically validated. The cryptographic layer (fuzzy extractor, ZKP) is scaffolded with placeholders.

The whitepaper reports a negative result: under standard fuzzy extractor bounds, the error rate required for 2D face recognition in the wild forces the helper string to leak more min-entropy than a 2D face embedding is estimated to contain. Secure 256-bit key extraction from a 2D camera is not achievable under these assumptions.

ZKHD/
├── Code/
│   ├── hv.py                 SimHash encoder, projection matrix, integrity guard
│   ├── simulate.py           Full pipeline: extract, encode, enroll, evaluate
│   ├── fuzzy_extractor.py    Fuzzy extractor (oracle mock)
│   ├── zk.py                 ZKP (Ed25519 mock)
│   └── requirements.txt
└── Dataset/                  Not tracked - place your datasets here

Dataset layout: Dataset/<name>/<user>/*.jpg/jpeg or flat Dataset/<name>/<identity>_*.jpg/jpeg.

cd Code

pip3 install -r requirements.txt

python3 simulate.py --enroll 3

The program prompts for backend (buffalo_l / antelopev2), dataset, and core count.