Biometric deduplication

In short

Biometric deduplication is the comparison of the face of the person signing up against all of an organization's previous biometric records, to detect whether that person is already registered under another identity. It doesn't check that someone is who they claim to be: it checks that they aren't someone who was already there. It's the only onboarding control that looks backward, at what the organization has already verified, instead of only looking at the case in front of it.

Comparing one-to-one and comparing one-against-many are different operations

Confusing the two makes people expect something from a verification system that it isn't doing.

• Verification, one-to-one — the captured face is compared against a specific reference, almost always the photo on the document. The question is whether those two images are of the same person.
• Deduplication, one-against-many — the captured face is compared against the entire database of previous records. The question is whether that person appears anywhere in that database.

The second is far more demanding, and the reason is arithmetic. Each comparison has a small probability of confusing two different people, and a search against the entire database runs millions of comparisons. With a fixed threshold, the larger the database, the more false matches appear. A system calibrated to verify and used to deduplicate returns a list of look-alikes that someone has to review by hand.

That's why deduplication is designed differently: its own thresholds, filters that narrow the search space before comparing, and a human review queue for what falls in the ambiguous zone. It's a process, not a single call.

What a deduplication finds

Not everything that turns up is fraud, and treating it as if it were creates friction for legitimate customers.

• The same person under two identities — the case that justifies the control. Different documents, different data, the same face.
• The return of a blocked account — someone who was rejected or removed, trying again with different data. Without deduplication, the block only lasts until the next sign-up.
• Incentive abuse — multiple accounts from the same person to capture bonuses, promotions or per-user limits. It's an economic problem before it's a regulatory one, and it's usually the first one to show up.
• The honest duplicate — a person who signed up twice without meaning to, because they didn't remember their earlier account. It's common and shouldn't be treated as an attempted fraud.

The distinction between the last two and the first two isn't made by the algorithm. It's made by the policy the organization defines for the result.

It's the control that exposes synthetic identity

A synthetic identity is built by combining real data from different people, or real data with invented data, until it forms an identity that belongs to no one but holds up against any individual check.

That's the problem: every onboarding control lets it through for good reasons. The document can be consistent. The data can exist. The face can belong to a real person, because it usually does. Nothing an isolated verification looks at has grounds to object.

What gives it away is repetition. The same person shows up behind several constructed identities, and that coincidence is only visible by comparing against the full history. A control that evaluates each sign-up separately can't see the pattern, because the pattern isn't in any single sign-up.

The same logic applies against a reused deepfake: a synthetic face used across several sign-ups stops being unique after the second one.

What implementing it requires

Deduplication isn't just another check you turn on. It changes what the organization stores and takes on new obligations.

• One biometric template per person — the numeric representation of the face that gets compared. It's what's stored and searched, not the photograph.
• A historical database — the control only works against what's already registered. A freshly created database finds nothing, and the value builds up over time.
• Sensitive personal data treated as such — biometrics carry a reinforced category under the region's data protection regimes. It implies a legal basis for processing it, clear disclosure to the person, retention periods and control over who accesses it. It's a conversation with the legal team before it's one with the vendor. The applicable frameworks by country are on the Colombia and Mexico pages.
• A policy on the result — what to do with a match: block, merge records, route to review. Without that definition, the system produces alerts nobody resolves.

How VU approaches it

Biometric deduplication rests on two VU capabilities that work on the same capture. Verify resolves the sign-up and produces the person's biometric template; Protect against fraud is the real-time detection capability that acts on the result of that comparison.

The one-against-many comparison is only as good as the capture is reliable. If the face entering the database can be a photo or a played-back video, the database fills with references that belong to no one. That's why liveness detection is a precondition, not an add-on: VU's is certified by iBeta at Level 2 of its testing program, which applies the methodology of the ISO/IEC 30107-3 standard to evaluate presentation attack detection.