Not a photo, not a fingerprint image, a string of numbers that represents a face well enough to compare it to another.
(Originally published on the FaceTrust AI Blog)
A numeric representation, not an image
A face embedding is a fixed-length vector of numbers, often several hundred dimensions, produced by a neural network from a face image. It's not a compressed photo; it's a mathematical representation of the facial features the network was trained to distinguish.
Why that representation is useful
Two embeddings generated from photos of the same person end up numerically close to each other, while embeddings from different people end up further apart. Comparing two faces becomes a matter of measuring the distance between their embeddings.
How this works in production (Cosine Similarity & 1:N Search)
In practice, modern facial recognition pipelines extract 68+ geometric facial landmarks to align a face, then pass it through a neural network (such as ArcFace) to generate a 512-dimensional vector.
Instead of comparing pixels or running slow text queries, your backend compares two vectors using cosine similarity:
- 1:1 Verification: Checking if a live capture matches a single enrolled user (e.g., confirming a similarity score ≥ 80%).
- 1:N Identification: Searching a live face embedding against an entire vector index of enrolled users in milliseconds.
A privacy advantage, not just a technical one
Because an embedding isn't a photo, storing one instead of a photo library limits what's exposed if the data is ever compromised, the embedding doesn't straightforwardly reconstruct back into a visible face image.
At FaceTrust AI, we discard raw registry photos right after initial NIN/BVN verification and store only a single canonical 512-d embedding per user.
Want to dive deeper into biometric architecture or test our identity verification API? Check out facetrustai.com or read more on Face Verification.
