At a glance
- Who should read this
- Fraud, identity and onboarding teams in financial services and any organization that verifies identity remotely.
- The risk
- Generative AI makes it easier to produce realistic faces, forged documents and live-looking video, and to run many consistent synthetic identities at once. Single checks, such as a selfie match, can be passed by fully fabricated inputs.
- Priority actions
- Layer independent checks so that no single check decides.
- Ask vendors how they detect injection attacks, not only presentation attacks.
- Link applications to find shared attributes, and keep monitoring after onboarding.
Summary
A synthetic identity is an invented person, usually built from a mix of real and fabricated details. Unlike impersonation, there is often no real victim who will notice and complain, and no real person to call back. Synthetic identities are used to open accounts, build credit and then borrow as much as possible before disappearing, a pattern known as a bust-out. Losses are often recorded as unpaid debt rather than fraud.
Generative AI lowers the cost of creating the evidence these identities need. US authorities, including FinCEN, have warned financial institutions about deepfake images and documents used to get past identity verification. This briefing explains where remote onboarding checks are weak, and how to layer them.
What each check confirms
| Check | Confirms | Does not confirm on its own |
|---|---|---|
| Data checks | Details exist and are consistent | That the applicant is the person described |
| Document check | The document appears genuine | That a real person is holding it |
| Selfie match | The face matches the document photo | That either face belongs to a real person |
| Liveness detection | A live person is likely in front of the camera | That the camera feed was not injected |
| Device and capture signals | The capture appears to come from a real device | That the identity is real |
A generated face placed on a forged document can pass a selfie match, because the match only compares two images. This is why checks have to be layered.
Presentation and injection attacks
Remote biometric checks are attacked in two ways:
- Presentation attacks show something to the real camera, such as a printed photo, a video replayed on a screen, or a mask. Presentation attack detection, often loosely called liveness detection, is designed to catch these, and is tested using methods set out in ISO/IEC 30107-3.
- Injection attacks bypass the camera, replacing its feed with prepared or synthetic video, for example through virtual camera software or device emulators. Liveness detection alone may not catch them. Defense depends on checking the integrity of the device and the capture process, an area addressed by the European technical specification CEN/TS 18099.
Strong presentation attack detection does not imply injection defense. Ask vendors about both, and how each is tested. NIST SP 800-63A-4 also requires presentation attack detection tested to ISO/IEC 30107-3 for remote biometric collection, and sets controls for detecting injected and forged media.
Layered proofing
NIST SP 800-63A-4, published in 2025, describes identity proofing in three stages: resolution, which narrows a claim to one unique identity; validation, which confirms the evidence is genuine and accurate; and verification, which confirms the applicant is the person the evidence describes. Two further practices are especially effective against synthetic identities:
- Linking. Look across applications for shared attributes, such as the same device, phone number range, address pattern or email structure appearing on supposedly unrelated people. Synthetic identities are often produced in batches, and the batch is easier to see than any single case.
- Weighing. Combine all signals into one risk decision, with outcomes of approve, step up for stronger evidence, manual review or decline, rather than letting any single check decide.
Stronger evidence can include reading the cryptographically signed chip in an electronic passport or identity card, where supported, which is much harder to forge than an image of a document.
Monitoring after onboarding
A synthetic identity reveals itself over time. Monitor for new links to other accounts, rapid requests for higher limits, contact details changed soon after opening, unusual access patterns, payments from unrelated third parties, and balances suddenly approaching every limit. Feed what you find back into onboarding rules.
Treat genuine customers fairly
Many signals associated with synthetic identities, such as a thin credit file or a recent change of address, are also common among genuine young people and newcomers. A signal is a reason to look more closely, not a verdict. Use consistent, documented review criteria, choose the lightest outcome that manages the risk, and check outcomes for unintended bias.
References
- NIST SP 800-63A-4, Digital Identity Guidelines: Identity Proofing and Enrollment, NIST, 2025
- FinCEN Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions (FIN-2024-Alert004), FinCEN, 2024
- ISO/IEC 30107-3:2023, Biometric presentation attack detection, Part 3: Testing and reporting, ISO/IEC
- CEN/TS 18099:2024, Biometric data injection attack detection, CEN (listing via the Estonian Centre for Standardisation)
This briefing draws on public research, standards and official guidance. It is general guidance, not advice for a specific environment. Related training: Synthetic Media and Fraud Defense.