How Do Synthetic and Manipulated Documents Affect KYC, and How Do You Catch Them?

How do synthetic and manipulated documents affect KYC?

Synthetic and manipulated documents affect KYC at the first evidence check. Remote onboarding usually asks for a photo of an identity document, reads its data and compares its portrait with a selfie. If the document image is fabricated or altered well enough, every later step works from false evidence, and the institution ends up with a customer who does not exist or who is not the person named.

Three forms matter in practice. A fully generated document image is created from a template or a generative model, with invented or stolen data and a portrait that may itself be synthetic. A manipulated genuine document starts from a real card and changes one thing, most often the photo or a data field such as date of birth. A digital template is a reusable file that produces many plausible cards with different names, which is what makes document fraud scale.

The consequences reach beyond onboarding. FinCEN's analysis of Bank Secrecy Act reporting found accounts opened with identities suspected to be produced with generative AI being used to receive and launder proceeds from scams, check fraud, credit card fraud, authorised push payment fraud, loan fraud and unemployment fraud, and as funnel accounts.

Why generative AI made forged IDs cheaper and harder to spot

Generative tools removed the skill barrier. On 5 February 2024, 404 Media reported on OnlyFake, a site that claimed to use neural networks to produce realistic fake ID photos for $15 each and said its generators could produce up to 20,000 documents a day. The reporter used an ID from the site to pass identity verification at OKX, a cryptocurrency exchange.

The image in the report was staged to look like a phone photo of a physical card lying on a carpet, which is the kind of picture many verification flows ask for. The fake did not need to survive a hologram tilt or a chip read, only a check of a single photo.

FinCEN saw the same shift in suspicious activity reports. Its alert FIN-2024-Alert004, issued on 13 November 2024, says that beginning in 2023 and continuing into 2024 institutions increasingly described suspected deepfake media in fraud schemes, often involving criminals altering or creating fraudulent identity documents to get around identity verification. It also notes that institutions often found these documents by re-reviewing account opening files after other suspicious activity, which means many slipped through at onboarding.

What controls catch an AI-generated or edited ID document?

No single control catches every forged document, so institutions stack checks that fail in different ways. NIST SP 800-63A-4 sets the baseline for remote identity proofing: providers shall analyse all digital media submitted during proofing for indications of modification, manipulation, tampering or forgery, and should analyse media for signatures of generative AI and deepfake tools. That applies to evidence images as much as to selfies.

FinCEN's red flags give a practical checklist for analysts: a photo that is internally inconsistent or does not fit the customer's stated age, several identity documents that disagree with each other, and an identity photo that a reverse image search matches to an online gallery of AI-generated faces.

A layered document control set for remote KYC usually includes:

• Read the chip on passports and national ID cards that carry one, and compare chip data and portrait with the printed page.
• Verify document numbers and key fields against the issuer or an authoritative database where your jurisdiction allows it.
• Run document authenticity checks on template, fonts, microprint and security features, and ask for a short tilt video when holograms matter.
• Inspect file metadata and capture source, and reject screenshots or edited files where policy requires a live capture.
• Compare the document portrait with a live selfie, and run a deepfake check on that selfie so a synthetic face cannot match a synthetic portrait.
• Search for the same face, layout or file across applications to catch template reuse.

Where Neural Defend fits: NeuroVerify on the selfie, NeuroDetect for suspect documents

NeuroVerify does not analyse ID documents, and it is not a replacement for a document verification provider. Its job in a document-led journey is the face side: the selfie, liveness clip or video that the face-match step compares with the document portrait. An attacker who forges an ID with a synthetic or borrowed portrait still has to produce a matching live face, and that face is usually a deepfake, a face swap or an edited photo. NeuroVerify is built to flag it.

One NeuroVerify request runs six signals: Visual Safety Screening, Live Presence Verification, Face Manipulation Detection, AI Generative Model Analysis, AI Pixel-Level Manipulation and AI Content Compliance, and returns one unified risk score with signal-level evidence. Eligible image requests return in under 200 ms p95 under defined test conditions. On the public VendorBench-100 benchmark of 100 adversarial still images, Neural Defend ranked first of 36 systems with an MCC of 0.876 against 0.646 for the next vendor; one benchmark author also co-authored Neural Defend's ARCAS 1B paper, and the corpus is public.

NeuroDetect covers the investigation side. When a document is flagged by your document vendor, an analyst or a FinCEN-style re-review, the analyst can upload the authorised document image into a NeuroDetect case. The workspace keeps model and version context, shows explainability heatmaps and attention views down to regions, adds an LLM-assisted plain-language summary linked to the evidence, and produces a downloadable forensic report with observations, uncertainty and analyst notes. The analyst makes the final call.

A review path for a suspect identity document

When a document looks wrong, the order of steps decides whether the evidence holds up later. A short, repeatable path works better than ad hoc escalation.

Hold the application and preserve the original file exactly as received, with capture metadata, device and session data and timestamps. Pull the matching selfie and its NeuroVerify result, since a high face-manipulation score on the selfie strengthens the case against the document. Open a NeuroDetect case for the document image and record the analyst's observations and uncertainty. Check the other documents and data the customer supplied for consistency. Decide under policy, and in the United States reference the key term FIN-2024-DEEPFAKEFRAUD in any related suspicious activity report, as FinCEN requests.

Then look backwards. FinCEN's experience is that forged documents are often found on re-review, so search for the same portrait, layout or device across past applications. The synthetic document and selfie fraud guide covers how these attacks are paired.

What document and deepfake checks cannot prove

A clean document result does not prove that the applicant is the document holder, and a clean selfie result does not prove the document is genuine. Each check covers its own evidence. A genuine stolen document paired with a genuine accomplice is invisible to both media checks and must be caught by data, behaviour and network signals.

Forensic views in NeuroDetect show where a model sees anomalies; they are evidence for an analyst, and compression, resizing and photos taken of screens can produce artefacts in genuine files. Thresholds and review rules belong to your policy, and they should be tested against your own genuine traffic before go-live.