Seventh Sense
Whitepapers

SenseCrypt
Technical library

Types of biometric storage

From face image to disposable face tokens

Similar sign-in screens can hide very different stored data. This paper maps face images, templates and protected tokens, explains their breach implications, and places SenseCrypt's face token within that taxonomy.

5 sectionsPDF · 6 pagesSeventh Sense / SenseCrypt
On this page

01

Ask what the system stores

Ask any face verification vendor what their product does and the answer comes back the same: it confirms that this is the same person. The difference between products sits downstream of that answer, in what the system is still holding once it has been given, and that is where the language in a data sheet stops being precise. Encrypted, hashed, tokenized and anonymized can appear within a single paragraph, and nothing in the document obliges them to mean different things.

Whoever runs the assessment absorbs that imprecision. A review with no vocabulary for the difference ends up ranking adjectives, which is a contest the strongest architecture does not necessarily win. A review that has one can place any product on a map with a single questionnaire, and the placement stops depending on which term the vendor preferred.

The map is built to be used against us as readily as by us. Anything that only returns a flattering result when applied to its author is marketing with footnotes, so the questions here are the ones we would want a reviewer to take to an incumbent system, to a competitor, or to an internal service nobody has examined in three years.

Templates, embeddings and feature vectors

Template, embedding and feature vector describe a single stored object: a numeric representation produced by a model so that two captures of one person land close together. Treating the three as synonyms is the first move that makes a review comparable across vendors.

A compromised face cannot be replaced

A compromised password is replaced in seconds. A representation engineered to survive aging, glasses and changing light keeps working for as long as its owner keeps the face, which is what turns a leaked template from an incident into a lifetime identifier.

Stolen biometric data can retain its value

Password dumps decay as people rotate credentials. A store of face-derived records moves the other way: each new deployment able to compare against it adds to what the original theft was worth.

02

What the paper covers

One pass covers the whole spectrum, capture through token, and the section that follows reduces the exercise to a single question. It is the question that separates a product whose protection is a property of its construction from one whose protection is a description, and vendors either answer it precisely or change the subject.

ISO/IEC 24745:2022 is the anchor. The standard names the properties a protected biometric system should exhibit, and the schemes it covers fall into two families that are much further apart than procurement documents tend to assume. ISO/IEC 30136:2018 supplies the metrics for testing schemes in either family. Both are cited here as a reference frame, not as certifications held.

Our own position is stated once, in the same vocabulary as everything else on the map, and then examined by the same criteria: IdP retains quantum-safe, sealed, biometric-free, disposable face tokens and their verifier challenges. Standard OIDC/SAML account records, device public keys, sessions, logs and encrypted tenant signing keys also persist. No face images or templates are retained in the documented phone flow. Licensed on-premises Webcam processes captures inside the customer deployment.

  1. What the numbers inside a stored template can still do once they leave the system that produced them
  2. The one adversary encryption at rest genuinely defends against, and the follow-up question that matters after a vendor says the templates are encrypted
  3. The two families ISO/IEC 24745 recognizes, and why deciding which one a product belongs to settles most of an assessment
  4. For each artifact on the map: what a breach of it hands an attacker, and whether revocation is even available
  5. Where the face token lands, and why its four ISO-named properties fall out of the construction instead of being administered on top of it
  6. What the placement means for data governed under GDPR, BIPA and the CCPA, and where the engineering fact ends and counsel begins

03

Who should read it

The intended reader is a security reviewer or privacy counsel with a vendor assessment already in front of them, and a preference for a short list of exact questions over another stack of data sheets. Its closing checklist carries no assumption about whose architecture is on the table, which is what makes it usable against an incumbent, a rival or us.

It also suits an internal audience that has inherited a biometric system somebody else procured. Placing that existing store on the map usually takes an afternoon, and the result tends to reorder the rest of the risk register.

04

Apply the taxonomy to your assessment

Six pages, two diagrams and one comparison table, short enough to finish between meetings rather than file for a quieter week. The most productive order is to run the closing questions against whatever you already operate first, so the vendor conversation begins from a position you have already tested.

The PDF is hosted on the SenseCrypt site behind a request form. If a form is inconvenient, ask us and a copy will come by email. Our engineers would rather field the questions the paper provokes than the ones it already answers.

05

Frequently asked questions

Is the taxonomy specific to SenseCrypt?

Mostly not. The spectrum, the two families named in ISO/IEC 24745 and the breach comparison hold for anything that verifies a face, and a reviewer is meant to point them at an incumbent or a rival. Our own construction appears once, in the final section, graded against criteria that were established before it was introduced.

Do I need a cryptography background to follow it?

No. Constructions are explained before they are used. One section names primitives with the precision a cryptographer would want, and skipping that detail costs nothing in the argument built around it.

Does encryption at rest not solve this already?

It solves one problem: theft of the storage medium without the keys. A matcher cannot compare ciphertext, so the decryption key has to be online, next to the data, and used on every verification. The productive question is therefore not whether data is encrypted but who can decrypt it, and how routinely that happens by design.

Our incumbent vendor has a SOC 2 report. Does this add anything?

A SOC 2 report describes the controls an organization runs around its data. It does not describe what the data is, so a store of recoverable templates and a store of sealed tokens can sit behind attestations that read identically. The taxonomy asks the other question: what would the holder of that store be able to do if every one of those controls failed at once.

Next step

The next level of detail depends on your stack, so the fastest route is a conversation.