Deepfake
A face or voice generated with artificial intelligence to represent someone who isn't there. What the term specifically means, why the content isn't the same as the vector it arrives through, and where it shows up within an identity journey.
In short
A deepfake is audio or video content generated or altered with machine learning models to depict a person saying or doing something they didn't say or do. Applied to identity, it's a synthetic face that mimics a document holder with enough quality to pass a biometric comparison.
The term names the content, not the attack. It's the piece the attacker manufactures, and on its own it says nothing about how it reaches the system that will evaluate it.
The content and the vector aren't the same thing
It's the distinction most definitions skip, and the one that decides which control applies: the deepfake is the content; the presentation attack and the [injection attack](/glosario/ataque-de-inyeccion) are the two vectors through which that content reaches verification.
- Presentation attack — the synthetic face is shown to the device's real camera, on a screen, a print or a mask. The camera works normally and captures what's in front of it.
- Injection attack — the synthetic face replaces the video stream before it reaches the application, using a virtual camera or device manipulation. The image never passed through a lens.
The same file works for both vectors. The controls that stop them are different, which is why it's worth naming which of the two is under discussion before comparing vendors.
Where a deepfake shows up in an identity journey
A deepfake doesn't attack the impersonated person. It attacks the moments when an organization decides whether the person on the other side is who they claim to be.
- Account sign-up — the synthetic face is presented against the holder's document so the system registers a verified identity that was never actually there.
- Account recovery — the same face is used to claim an existing account, which is the route to an account takeover.
- Authorizing a transaction — when biometrics is what confirms a transaction, the synthetic face targets that confirmation.
In all three cases the biometric comparison correctly answers the question it's asked: whether the captured face looks like the one on the document. The question left unasked is whether that face belongs to a real person present at that moment.
What control answers a deepfake
Liveness detection is the control that determines whether there's a real person in front of the camera, and it's the answer to the presentation vector. Injection detection is a different control and covers the other vector.
The full attack sequence, step by step, with the control that corresponds to each one, is on the fraud type page: deepfake.
Frequently asked questions
It's audio or video content generated or altered with machine learning models to depict a person saying or doing something that didn't happen. In identity verification, it's a synthetic face that mimics a document holder with enough quality to pass a biometric comparison. What turned it into an operational problem isn't the technique, which has existed for years, but that producing one stopped requiring equipment or specialized knowledge.
The deepfake is the content and the presentation attack is one of the two ways to deliver it to the system. In a presentation attack the content is shown to the device's physical camera. In an injection attack the video stream is replaced before the application receives it, without going through the camera. A defense that only covers the first leaves the second open.
No, because it answers a different question. The biometric comparison measures the resemblance between the captured face and the one on the document, and a well-built synthetic face resembles it. The question that detects the attack is whether there's a real person present at the moment of capture, and liveness detection answers that one.
