KYC deepfakes: what a selfie check can and can't tell you

A selfie or video check answers narrower questions than most teams assume. A face match tells you the selfie and the ID photo probably show the same face. A liveness check tells you a photo, screen or mask probably wasn't held up to the camera. Neither tells you the ID is genuine, or that the video reaching your system came from a real camera rather than a generator. Deepfakes aim at those gaps, so pair the selfie check with document checks, defenses against injected video and a step-up when something looks wrong.

In November 2024, FinCEN warned banks about fraud schemes using deepfake media. Its alert describes criminals using generative AI to alter or create the images on driver's licenses and passports, and combining those images with stolen or fake personal details to build synthetic identities.

What each check tells you

• Document check. Whether the ID looks genuine and unedited. It doesn't tell you who is holding it.
• Face match. Whether the selfie and the ID photo show the same face. It doesn't tell you either image is real. A generated face on an edited ID can match a generated selfie perfectly.
• Liveness. Also called presentation attack detection, it checks for a photo, screen or mask held up to the camera. NIST's identity guidelines require it to be tested under ISO/IEC 30107-3, which covers attacks at the capture device. It doesn't tell you whether the video was swapped after the camera.
• Injection detection. It looks for forged media inserted between the camera and the check, for example from a virtual camera, a device emulator or a jailbroken phone. NIST says plainly that a biometric comparison on a captured sample does not prevent these attacks, and that even live capture and liveness checks are not enough on their own. A European technical specification, CEN/TS 18099, sets out how to test it.
• Media analysis. Whether the image or video itself shows signs of editing or generation. NIST's guidelines recommend analyzing media for signatures of generative AI and deepfake tools.

How deepfakes get in

• Generated or edited ID images, with a matching generated selfie.
• Synthetic identities: a generated face attached to real or invented personal details.
• Pre-made video played through a third-party webcam plugin or virtual camera, so the check sees generated video instead of a live camera. FinCEN names webcam plugins in its alert, and ENISA's good practices for remote identity proofing explain how a virtual camera replaces the physical one.
• Face-swapped or generated video shown to the camera or injected into the stream.

Red flags a reviewer can check

FinCEN's alert lists red flags. Several can be checked by hand:

• The photo is internally inconsistent, with visual tells of being altered. Look at the edges of the hair, ears and glasses, and at light and shadows that don't match the background.
• The applicant's identity documents are inconsistent with each other.
• A reverse image search of the ID photo or selfie matches an online gallery of AI-generated faces.
• The applicant uses a third-party webcam plugin during a live check, or the session has excessive or suspicious technical glitches.
• The applicant declines multifactor authentication.
• Location or device data doesn't fit the identity documents.

FinCEN also notes that no single red flag is necessarily a sign of wrongdoing. Look for more than one, and for a story that doesn't hold together.

Step up when something looks wrong

1. Ask for a different proof: another document, a fresh capture on another device, or a live session you start.
2. Make it interactive. In a Hong Kong case where a company lost about HK$200 million after a video meeting, the Hong Kong government's account says the meeting was pre-recorded and there was no real interaction. Unscripted questions raise the bar, but passing them is not proof on its own.
3. Write down what you saw and where, and escalate to your fraud or compliance team. US institutions filing a suspicious activity report on this kind of fraud are asked to include the key term FIN-2024-DEEPFAKEFRAUD.

Phone cameras smooth skin, compress video and apply filters, and that can look like editing. A flag is a reason to step up verification, not proof that the applicant isn't who they say.

Where software helps

A reviewer can look closely at one applicant. Onboarding runs in minutes and at volume, and traces of editing or generation can be too small to see at normal speed. Software can check every ID image, selfie and video at upload, look frame by frame for generation and editing traces, read camera data and file history, and mark where each finding sits. It cannot prove who someone is. That still takes your identity verification provider, your own records and a person's judgment.

FAQ

Can a liveness check stop deepfakes?

It is built to stop a photo, screen or mask held up to the camera. Video injected after the camera skips that point, so liveness needs injection detection and media analysis beside it.

What is an injection attack?

NIST defines it as an attack in which someone supplies untrusted biometric information or media into a program or process. In onboarding, that usually means feeding pre-recorded or generated video into the check, for example through a virtual camera.

What does FinCEN say about deepfakes in KYC?

Its November 2024 alert (FIN-2024-Alert004) describes generated and altered identity documents, synthetic identities and webcam plugins, lists red flags, and asks institutions to use the key term FIN-2024-DEEPFAKEFRAUD in suspicious activity reports.

Does a deepfake flag mean the applicant is a fraudster?

No. Filters, compression and poor lighting can look like editing. A flag means a closer look or another proof, with a person deciding.

Do we still need an identity verification provider?

Yes. Checking files for edits and generation adds to identity verification. It doesn't replace liveness, biometric matching or database checks.

How Trustable helps

Trustable checks the files your onboarding flow already collects. ID images are checked for edited photos and details, rebuilt templates, signs of AI generation and camera data that conflicts with how the image was supposedly captured. Selfies are checked for AI generation and editing, and recorded or live video for face swaps and other generation traces. Each finding is marked on the image or frame with its strength, for your compliance or fraud team to review. Trustable is not a liveness or biometric-matching product and does no database lookups; it works alongside the identity verification provider you already use. Reading Content Credentials is in development. See fake ID and deepfake detection for onboarding, or Book a demo.