AI-generated synthetic identity fraud defeats account opening not by stealing a real customer’s identity but by manufacturing a believable one from real and fabricated data, then dressing it in AI-generated documents, photos, or video convincing enough to clear standard identity verification. The defense is not a single detection score. It is a layered onboarding control that checks document authenticity, biometric liveness, data consistency across sources, and behavioral signal together, because a fraud kit that defeats one layer rarely defeats all four. This guide gives a BSA/AML officer, head of fraud, or compliance leader at a bank or fintech a framework for assessing that layered control and for governing the AI-based identity-verification tools doing the checking.
This is for the leader who already treats synthetic identity and deepfake document fraud as a named risk and needs a structured way to assess whether the institution’s onboarding stack actually catches it, not another explainer of how generative AI creates fake IDs. It assumes an existing Customer Identification Program (CIP) and Customer Due Diligence (CDD) process under the Bank Secrecy Act; the gap this guide addresses is where that process assumes a 2019-era fraud attempt and an AI-assisted applicant has already moved past it. DSE runs this review as part of a fixed-fee AI security assessment; the framework below applies whether you run it internally or bring in a third party.
Want to put this into practice? Use our AI vendor due-diligence questionnaire.
Why onboarding fraud is a different problem than payment fraud
A separate, well-covered category of AI fraud targets the approval moment on an existing account: a cloned voice or deepfake video impersonating a CFO or executive to push through a wire transfer. DSE’s AI voice cloning fraud risk assessment covers that threat and the out-of-band callback control that stops it. Synthetic identity and deepfake onboarding fraud is a different attack surface entirely. It does not target a trusted person’s authority over an existing account; it targets the moment a new identity is created in the first place, before there is any account history, any established behavioral baseline, or any relationship to lean on. The control point is not “does this match who we already trust,” it is “does this identity, and the person presenting it, actually correspond to each other and to a real individual.”
That distinction matters because the two fraud types need different controls and sit with different owners. Payment fraud is a treasury-operations and approval-workflow problem. Onboarding fraud is a CIP, CDD, and identity-verification-vendor problem, and it is where most institutions’ AI governance attention has not yet caught up with their fraud exposure.
The three techniques defeating standard identity checks
Generative AI has made three onboarding fraud techniques materially cheaper and more convincing in the last two years, and each one targets a different part of a standard identity-verification flow.
| Technique | What it does | Which check it targets |
|---|---|---|
| Synthetic identity construction | Blends a real Social Security number fragment, a fabricated name, and a plausible address history into a persona with no real underlying individual | Database and bureau checks that confirm an identity exists but not that it belongs to a single real person |
| Document fabrication and manipulation | Uses generative tools to create a convincing fake driver’s license, passport, or utility bill, or to alter a real document’s photo and data fields | Document-authenticity checks that rely on visual inspection or template matching rather than forensic signal analysis |
| Presentation and injection attacks on liveness checks | Presents a deepfake video, a digitally injected image feed, or a 3D mask to a selfie-match or liveness step instead of a live human face | Biometric liveness detection and selfie-to-document face matching |
A fraud ring rarely relies on just one of these. A typical attempt pairs a synthetic identity (so the underlying data passes a bureau check) with a fabricated or altered document and a liveness-evading presentation attack, because stacking the techniques is what gets the application through an onboarding flow built to catch any one of them in isolation.
A five-layer onboarding identity assurance framework
The control that holds against this combination is not a single vendor score. It is five layers that each catch a different part of the attack, so a fraud attempt that clears one layer still has four more to defeat.
| Layer | What it checks | Common failure mode when missing |
|---|---|---|
| Document forensics | Pixel-level and metadata analysis of the submitted ID for signs of AI generation or manipulation, not just template and hologram matching | Visual-only checks pass a well-made fabricated document because it looks correct to a human reviewer and a template matcher |
| Biometric liveness and injection detection | Confirms the selfie or video came from a live, physically present human and was not digitally injected, replayed, or synthetically generated | Liveness checks built before deepfake-grade video existed test for blinking and motion, which a modern presentation attack now reproduces |
| Cross-source data consistency | Compares the applicant’s claimed identity data against multiple independent sources (bureau, device, carrier, address history) for internal contradictions | A single-bureau check confirms the identity exists without checking whether every data point about it agrees with every other data point |
| Device and behavioral signal | Evaluates the device, network, and session behavior the application arrived through, including signals common to automated or bulk fraud attempts | Fraud rings submit dozens of synthetic identities through the same device fingerprint or network pattern, invisible to an identity-only check |
| Ongoing monitoring past onboarding | Re-screens the account’s early activity for patterns consistent with a synthetic identity that passed onboarding but behaves differently than a genuine new customer | A model validated once at onboarding has no process for catching an identity that slipped through until the account is already active |
Most institutions have built layers one and two reasonably well, because that is what vendor marketing for identity-verification platforms emphasizes. The fourth and fifth layers are where DSE finds the most exposure, because they require connecting onboarding fraud data to fraud and AML systems that were built and are owned separately.
Where this becomes an AI governance problem, not just a fraud problem
The identity-verification tool doing this screening is itself an AI system, usually bought from a vendor, and it needs the same governance scrutiny any model-driven decision in a bank requires. Three governance obligations apply whether or not the institution built the tool itself:
Vendor AI risk review. An identity-verification vendor is a third-party AI vendor like any other, and the June 2023 interagency third-party guidance’s lifecycle applies to it: due diligence on how the vendor’s model was trained and tested against deepfake-grade attacks, contract terms requiring notice when the vendor updates its detection model, and ongoing monitoring of the vendor’s actual catch rate rather than its marketing claim. DSE’s machine learning third-party risk management guide applies that full lifecycle to a model-supplying vendor; an identity-verification platform is exactly this kind of vendor.
Testing under a recognized framework. The NIST AI RMF’s MEASURE function calls for testing an AI system’s performance against the conditions it will actually face, which for an identity-verification tool means testing its detection rate against current-generation deepfake and document-fabrication samples, not only the synthetic test set the vendor supplies. A tool’s accuracy claim from eighteen months ago says little about its performance against today’s generation tools.
Disparate impact in verification friction. Biometric liveness and face-matching systems have documented accuracy differences across demographic groups, which means a higher false-rejection rate for some applicants is a plausible outcome of the tool itself, not only of fraud attempts. An institution relying on this kind of system carries an obligation, independent of the fraud question, to test whether legitimate applicants are being disproportionately rejected or burdened by additional verification steps along lines that intersect with ECOA and Regulation B.
None of this replaces the institution’s existing CIP, CDD, and GLBA safeguards obligations for the nonpublic personal information these tools process. It adds a documented basis for the governance and vendor-oversight questions an examiner or an internal audit will ask once an AI system, not a person, is making the first pass on whether an applicant is who they claim to be.
A compliance evidence checklist
Pull these together and a BSA/AML or compliance function has a defensible record of how onboarding identity assurance works and who owns each piece:
- [ ] Documentation of all five assurance layers in the current onboarding flow, with the vendor or system responsible for each
- [ ] Vendor due-diligence records showing how the identity-verification vendor tests its model against current deepfake and document-fabrication techniques
- [ ] A contract clause requiring vendor notice when the underlying detection model changes or is retrained
- [ ] Test results measuring the tool’s actual catch rate on recent-generation synthetic samples, not only the vendor’s own benchmark
- [ ] A disparate-impact review of false-rejection and step-up-verification rates across demographic groups
- [ ] A process connecting onboarding fraud signals to post-onboarding account monitoring, so an identity that clears onboarding is still screened
- [ ] Named ownership across fraud, BSA/AML, and model risk functions for the identity-verification tool, so no single team assumes another team is covering the AI governance piece
Questions buyers ask
Is synthetic identity fraud the same risk as the deepfake voice fraud used in wire transfers?
No. Synthetic identity and deepfake onboarding fraud targets the creation of a new account and new identity, defeating CIP and CDD checks before any account history exists. Deepfake voice and video fraud targets an existing trusted relationship, impersonating a known executive or customer to push through a payment approval. They need different controls and usually sit with different owners: fraud and BSA/AML for onboarding, treasury operations and approval workflow for payment fraud.
Can a single identity-verification vendor’s product catch all of this on its own?
Rarely on a sustained basis. Most identity-verification platforms cover document checks and liveness detection well, which is the first two of the five layers. The techniques that defeat a single vendor’s product are the ones that stack a synthetic identity with a fabricated document and a liveness-evading presentation attack, which is why cross-source data consistency, device and behavioral signal, and post-onboarding monitoring need to sit alongside the vendor’s tool rather than be assumed inside it.
Does the BSA/AML program already cover this, or is this a separate AI governance workstream?
Both, and that is the gap. The BSA/AML program owns the CIP and CDD requirements this fraud attacks. The AI governance workstream owns the vendor oversight, testing, and bias review of the AI tool doing the screening. A program that treats the identity-verification vendor purely as a BSA/AML tool and never subjects it to AI vendor risk review and NIST AI RMF-aligned testing has a gap an examiner reviewing AI governance maturity is likely to find.
How often should the identity-verification tool be retested against new deepfake techniques?
At minimum annually, and immediately after any material update to the vendor’s detection model or any observed increase in successful onboarding fraud attempts. Generative AI fraud tools improve faster than most vendor benchmark refresh cycles, so a test result from a year ago understates current exposure more than a similar gap would for most other AI systems.
Can DSE certify that our onboarding identity-verification program meets BSA/AML and AI governance expectations?
No. DSE prepares programs for audit and examination by documenting the five-layer control, the vendor due-diligence record, and the testing and bias-review evidence an examiner or auditor expects, and the result is audit-ready and supervisory-framework-aligned. DSE does not certify compliance or guarantee any examination outcome; that determination rests with your examiners and your own BSA/AML and risk functions.
The Bottom Line
AI-generated synthetic identity fraud at onboarding is not solved by a stronger version of a single identity-verification check. It is solved by five layers working together, document forensics, biometric liveness, cross-source data consistency, device and behavioral signal, and post-onboarding monitoring, so a fraud attempt that defeats one layer still meets four more. The identity-verification tool running the first two layers is itself an AI system that needs vendor risk review, testing against current deepfake techniques, and a bias review of its rejection rates, not just a line item in the BSA/AML program.
Building that five-layer control and the AI governance record behind the identity-verification vendor is exactly what DSE’s AI security assessment is built to deliver, with AI vendor risk review for the identity-verification vendor specifically. Start with the AI governance checklist to see where your onboarding controls stand today, or review finserv compliance for the broader financial-services compliance picture.
Key facts
- FinCEN's November 2024 alert on deepfake media warned that generative AI tools now let fraud rings fabricate or manipulate identity documents, photos, and videos well enough to defeat many identity-verification checks used at account opening (FinCEN, 2024).
- A synthetic identity combines real and fabricated identifying information into a persona that corresponds to no actual person, which lets it pass a single-point identity check that only confirms an identity exists in a data source rather than confirming the applicant in front of the camera owns it (DSE, 2026).