Picture an applicant who passes a selfie video prompt while the session runs through a third party webcam plugin (illustrative). The date of birth says 61; the face looks about 30. Two weeks later the new account sends a burst of payments to a digital asset exchange. Each step is a red flag in FinCEN's deepfake alert. None alone proves suspicious activity.
Deepfake detection is the layer of fraud detection and prevention built to catch those steps.
What is deepfake detection, and where does a financial institution need it?
Deepfake detection is the set of checks that decides whether a face, voice, document or video is genuine and captured live, or was generated or altered by AI. The EU's AI Act defines the term. Under Article 3(60), a deep fake is AI made or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and "would falsely appear to a person to be authentic or truthful".
A bank or payment firm meets deepfakes at the three points FinCEN's identity analysis uses: "account opening, account access, and during transactions". At opening, a generated face or document creates the customer. At access, a cloned voice or face passes as the real one, and in a transaction a call tells someone to send money.
FinCEN has seen a rise in suspicious activity reports describing suspected deepfake media, "beginning in 2023 and continuing in 2024".
How do deepfakes get through remote onboarding?
"A deepfake video is used to impersonate an individual portrayed on a stolen driver's license." NIST SP 800-63A-4, the US federal identity proofing guideline of July 2025, files that example under injection attacks.
FinCEN found criminals using generative AI to make "falsified documents, photographs, and videos" to get around customer identification and due diligence controls. Each fake aims at a different check: a forged ID card at the review that document fraud detection covers, a swapped face at the selfie match and liveness test of eKYC.
The fake travels one of two ways. A presentation attack is shown to the capture device itself. An injection attack skips the camera and feeds the fake into the data stream.
On live video checks, FinCEN notes applicants who claim repeated technical glitches or ask to switch channel mid check. NIST tells agents to look for "high latency, synchronization issues". It also asks for random cues, such as a request to move. Complaints in the IC3's 2025 report describe online job interviews where lip movement on camera does not fully match the audio.
What is an injection attack, and why can a liveness check miss it?
In NIST's terms, an injection attack slips forged media in between the capture point and the comparison. ETSI's proofing standard TS 119 461 gives the deepfake version: a recorded or generated stream "purporting to come from the camera". A liveness test judges what reaches it. A clean injected stream can look live. NIST says presentation attack detection is "not sufficient to address all possible cases" and requires checks that the media came from a genuine sensor, not a virtual camera or an emulator.
The selfie and liveness check stay with the firm's IDV provider, but an emulator, one of NIST's examples, also shows at the device layer. complead, formerly Sanction Scanner, reads device and session signals from before login to payment, flags each spoofing tool, emulator, tampered runtime or automation tool on its own, and gathers no face or voice data.
Which standards test liveness and injection attack detection?
None of these is a bank rule; each is a benchmark to put to the IDV provider.
Table 1. Liveness and injection standards (as of November 2026)
| Standard | What it covers | What it leaves out | Who it is written for | Status |
|---|---|---|---|---|
| ISO/IEC 30107-3:2023 | Testing and reporting of presentation attack detection (PAD) | Attacks away from the capture device, injection included | Any adopter; NIST requires it for PAD tests | Edition 2, January 2023 |
| NIST SP 800-63A-4 | Remote PAD with an impostor attack presentation accept rate (IAPAR) below 0.07; genuine sensor checks; media analysis; false negative rates on request | Authentication after enrollment, which 63B covers | Federal agencies, proofing providers and organizations using them | Final, July 2025 |
| NIST SP 800-63B-4 | PAD required for face login; voice comparison not allowed | Onboarding, which 63A covers | Login to US government systems | Final, July 2025 |
| ETSI TS 119 461 V2.1.1, citing CEN/TS 18099 | Detection of generated or manipulated faces in video; injection detection tested by an accredited lab | Proofing outside trust services | Trust service providers and their proofing subcontractors | Lab test before the end of 2026 |
| AI Act Article 50 | Providers mark AI output; deployers disclose deep fakes | Checks by the firm that receives a fake | Providers and deployers of AI systems in the EU | Applies from 2 August 2026 |
The attacks ISO/IEC 30107-3 considers "take place at the biometric capture device during presentation", and any other attack is "outside the scope", so a PAD test report to that standard says nothing about injected video. ETSI covers that for trust services: when video is captured on the applicant's own device, injection detection must pass an accredited lab test to TS 18099 before the end of 2026, at level 2 for Baseline proofing and level 3 for Extended, then again at least every second year.
What should a bank ask its identity verification provider?
- A PAD test report to ISO/IEC 30107-3. Ask which attack types and which lab; the report covers the camera only.
- An injection attack test to CEN/TS 18099. ETSI asks it of trust services for own device video before the end of 2026.
- False negative rates by attack type. NIST has providers make them "available to RPs upon request", and a bank buying the check sits in the relying party's seat.
- Sensor and device checks. Virtual cameras, emulators and jailbroken phones are NIST's own examples.
- How results reach the risk model. Ask which flags and scores the provider sends back with each result, not only pass or fail.
- Where the face data goes. A face image processed to identify someone is biometric data under the GDPR, and Article 9 bars that processing unless an exception applies, such as explicit consent.
What red flags did FinCEN list in its deepfake alert?
Table 2 sorts FinCEN's nine red flags by where each one shows and who decides.
Table 2. FinCEN's nine deepfake red flags (FIN-2024-Alert004)
| # | Red flag (FinCEN's words, shortened) | Where it shows | What checks it | Who decides |
|---|---|---|---|---|
| 1 | Photo looks altered, or does not fit the date of birth or other details | IDV session | Face match, document review | Onboarding analyst |
| 2 | Identity documents that contradict each other | IDV session | Document review | Onboarding analyst |
| 3 | Third party webcam plugin, or a channel change blamed on glitches, in live verification | IDV session | Liveness and injection detection | Provider flags, analyst decides |
| 4 | Customer declines multifactor authentication | Authentication | Login controls | Fraud team |
| 5 | Reverse image or open source search finds the photo in a gallery of AI made faces | IDV session | Reverse image search | Onboarding analyst |
| 6 | Deepfake detection software flags the photo or video | IDV session | Detection software | Provider flags, analyst decides |
| 7 | Software flags AI written text in the profile or prompt answers | IDV session | Text detection | Onboarding analyst |
| 8 | Location or device data that do not fit the identity documents | Device | Device and location signals | Fraud team |
| 9 | New or quiet account: rapid transactions, high volumes to risky payees, many chargebacks or rejected payments | Account activity | Transaction monitoring | Fraud or AML analyst |
FinCEN says no single red flag "is necessarily indicative" of suspicious activity. The firm weighs the surrounding facts first. When it files a suspicious activity report, FinCEN wants FIN-2024-DEEPFAKEFRAUD both in field 2 of the SAR and in its narrative, beside the key terms of the underlying scheme. A deceived customer in the US files a separate report of fraud with the FBI's IC3, as the FBI's announcement of 3 December 2024 asks.
How are deepfake voices and videos used against payments?
Against payments, a deepfake plays someone the victim already trusts. FinCEN describes family emergency schemes, where a cloned voice or video poses as a relative or friend, and executive impersonation, where a fake executive tells staff to send large sums to accounts the criminal controls. The same FBI announcement adds that criminals get into bank accounts with AI made audio clips of the holder, and stage real time video chats with "alleged company executives".
Picture a treasury analyst who receives a voice note from the finance director asking for an urgent payment to a new supplier (illustrative). The voice was cloned from a recorded talk. The request is CEO fraud with a synthetic voice, one form of impersonation fraud.
How do you verify a payment request made on a voice or video call?
- Hang up and call back. The FBI advises calling the organization's own number; use one the firm already holds, never the number in the message.
- A second approver. A payment asked for by voice or video waits for a second person's sign off.
- A code word. The FBI tells families to agree "a secret word or phrase"; an executive team can do the same for payments.
- No voice match as proof. NIST bars voice comparison for authentication, which sets voice apart from other biometric authentication methods.
- A payee check. When the request names a new beneficiary, look at that account before the money goes; complead checks that payee for mule links, reuse and known fraud signals.
Which controls catch deepfake fraud after the account is open?
Back to the opening applicant (illustrative): the selfie passed, but the phone behind it opened four other accounts that week, and logins come from an IP address far from the stated home. Device fingerprinting ties those accounts together. Two of FinCEN's indicators for further due diligence fit here: access from an IP address at odds with the customer's profile, and coordinated activity across similar accounts. On complead, the device and account graph surfaces "onboarding clusters from one device", spots one device used across several accounts and picks up impossible travel.
In FinCEN's data, firms tend to catch fake ID documents on a second pass over the opening file, and a reverse image search can tie an ID photo to a public collection of AI made faces.
FinCEN reports that such accounts also receive and launder the proceeds of other scams, and some serve as funnel accounts. AI fraud detection helps here when each score names the signals behind it, so an analyst can see why an account was stopped.
Which 2026 dates matter for deepfake controls?
- April. The IC3's 2025 report counted 22,364 complaints that mention AI, with $893,346,472 in adjusted losses: complaints, not prevalence.
- July. FATF's update on virtual assets said AI abuse is "emerging as a risk factor" in fraud, hacking and money laundering, and asked supervisors of virtual asset firms to take account of "controls against deepfakes and synthetic identities".
- 24 July. A FinCEN alert on federal student aid fraud warned that fraudsters may use AI to "overcome identity verification" with generated documents.
- 2 August. AI Act Article 50 began to apply; providers with systems on the market before then have until 2 December 2026 to mark output (Regulation (EU) 2026/1744).
- End of 2026. ETSI's last date for lab tested injection detection on own device video in trust services.
When a deepfake opens an account, is a real person's identity behind the face, or one stitched together from stolen and invented details? FinCEN's student aid alert calls the second a synthetic identity, and synthetic identity fraud has checks of its own.