This document describes the security properties, threat model, and limitations of BioProt.

1. Biometric Templates: Protected representations of face embeddings
2. User Seeds: Cryptographic seeds used for transform generation
3. KMS Storage: Encrypted file containing all user keys

We assume an adversary who may:

1. Obtain template database: Access to stored protected templates
2. Know the algorithm: Full knowledge of Ortho+Sign and Perm+LUT transforms
3. Have computing resources: Can perform brute-force and ML-based attacks
4. NOT have KMS access: Cannot decrypt the key management storage

Definition: Given a protected template T = f(x, s), it should be computationally infeasible to recover the original embedding x.

Implementation:

• Ortho+Sign: Sign binarization destroys magnitude information
• Perm+LUT: Quantization loses precision, LUT is one-way

Verification: See evaluate.py::naive_inversion_test() and regressor_inversion_test()

Definition: If a template is compromised, the system can issue a new template that:

• Does not match the old template
• Works correctly for legitimate authentication

Implementation:

• Key rotation generates new random seed
• New seed produces uncorrelated template
• Old template becomes invalid (random similarity ~0.5)

Verification: See evaluate.py::test_revocation()

Definition: Templates generated with different seeds should be statistically independent, preventing cross-database linking.

Implementation:

• Each user has unique seed
• Same embedding + different seeds → uncorrelated templates
• No pattern leakage between templates

Analysis:

For seeds s₁ ≠ s₂ and embedding x:
- T₁ = f(x, s₁)
- T₂ = f(x, s₂)
- E[similarity(T₁, T₂)] ≈ 0.5 (random chance)
- Var[similarity(T₁, T₂)] → 0 as dim → ∞

Definition: Protected templates should maintain recognition accuracy comparable to raw embeddings.

Implementation:

• Ortho+Sign: Orthonormal projection approximately preserves distances
• Both methods: Proper threshold calibration

Metrics:

• AUC should be > 0.99 for good systems
• TAR @ FAR=0.1% should be > 95%

master_key = PBKDF2(
    password=user_passphrase,
    salt=random_16_bytes,
    iterations=100000,
    hash=SHA256
)

• Algorithm: Fernet (AES-128-CBC + HMAC-SHA256)
• Each user's seed encrypted separately
• Integrity protected (HMAC)

{
    "user_id": {
        "encrypted_seed": "base64...",
        "version": 1,
        "created_at": "2024-01-01T00:00:00Z",
        "rotated_at": null
    }
}

This is a research prototype demonstrating concepts. For production:

1. Use Hardware Security Module (HSM) for key storage
2. Implement proper access control
3. Add audit logging
4. Use secure enclaves for matching

1. Local KMS: File-based storage is not as secure as HSM
2. Single passphrase: All users share KMS encryption key
3. No liveness detection: Assumes genuine face presentation
4. Fixed embedding model: Tied to 512-dimensional FaceNet

1. HSM/TPM integration
2. Per-user encryption keys
3. Distributed key management
4. Template update without re-enrollment
5. Secure multi-party matching

• Biometric data is "special category" under GDPR
• Templates may still be considered biometric data
• Key rotation supports right to erasure
• Irreversibility supports data minimization

This implementation aligns with ISO/IEC 24745 principles:

• Irreversibility (Section 5.2)
• Renewability/Cancelability (Section 5.3)
• Unlinkability (Section 5.4)

1. ISO/IEC 24745:2011 "Biometric template protection"
2. NIST SP 800-76-2 "Biometric Specifications for PIV"
3. Article 9 GDPR "Processing of special categories of data"