F

Architecture & Trust

Security & Isolation

Designed from first principles for NDPA and GDPR biometric data minimization, strict multi-tenant isolation, and explainable audits.

Template-first minimization

By default, FaceCore computes normalized facial embeddings in memory and discards raw images immediately after inference. Opt-in retention requires an explicit retention limit.

Envelope encryption at rest

Stored templates are encrypted per tenant using envelope keys backed by AWS KMS, Azure Key Vault, Oracle Vault, or configured local KMS wrapping.

PostgreSQL Row-Level Security (RLS)

Every gallery, subject, template, consent, and decision row carries tenant_id and is guarded by database-layer RLS policies. Cross-tenant queries return zero rows.

Consent-bound enrollment & hard deletion

Enrolling a subject requires a valid consent_id. Calling DELETE /v1/subjects/{id} or withdrawing consent purges templates from both relational storage and the vector index while keeping an immutable, non-biometric audit entry.

Replay-proof liveness manifests

Active and multi-view sessions bind randomized challenges to a single-use cryptographic nonce and verify an HMAC-SHA256 signature over frame hashes and capture timestamps.

Commercial model licensing gate

On startup and /v1/ready checks, FaceCore verifies every active model in the registry carries an approved commercial license before serving traffic.

Auditability Without Raw Image Exposure

Every decision log retains decision_id, request_id, model_versions, policy_version, calibrated threshold, and reasons[] — allowing compliance teams to reconstruct why a user passed or failed months later without storing their face photo.

Sample immutable decision record
{
  "decision_id": "dec_01j9m2x9m0",
  "tenant_id": "tnt_01j9k4m8p2",
  "operation": "verify",
  "result": "MATCH",
  "score": 0.614,
  "threshold": 0.452,
  "profile": "standard",
  "model_versions": {
    "detector": "scrfd-2.5g",
    "embedder": "arcface-r100-v1@1.2.0",
    "liveness": "minifasnet-v2"
  },
  "policy_version": "p-2026-01",
  "reasons": []
}