Banking, financial services & insurance
The customer passed liveness. The CFO approved on video. Neither was real.
Deepfakes now open bank accounts, authorise payments on video calls and pass voice checks in the contact centre. NeuroVerify checks the face, video or voice at each of those moments, NeuroDetect gives your investigators forensic evidence, and your policy makes the call.
The problem, in numbers
Hong Kong, Brussels, Washington: the same attack, priced in different currencies.
- US$25.6m
sent in 15 transfers by a finance employee in Hong Kong after a video call on which the CFO and every colleague were deepfakes. Arup later confirmed it was the victim.
CNN, citing Hong Kong police, 16 May 2024 - 1 in 5
biometric fraud attempts in identity verification were deepfakes. Deepfaked selfies rose 58% in 2025 and injection attacks 40% year on year.
Entrust 2026 Identity Fraud Report, 18 Nov 2025 - €4.2bn
of payment fraud across the EEA in 2024. Manipulating the payer caused more than half of credit-transfer fraud by value, and customers bore about 85% of those losses.
EBA & ECB Report on Payment Fraud, Dec 2025 - US$40bn
in US fraud losses enabled by generative AI by 2027, up from US$12.3bn in 2023. A projection, not a count, from Deloitte’s Center for Financial Services.
Deloitte Insights, 29 May 2024
Where deepfakes hit BFSI
Six moments where a bank decides on a face or a voice.
Each one was designed for a world where seeing and hearing a person was evidence. The attack is the same in all six: supply the face or the voice the control expects.
Remote account opening and video KYC
A face swap built from a stolen ID photo is fed through a virtual camera, so face match and liveness both pass. FinCEN reports accounts opened this way being used as funnel accounts for other frauds.
NeuroVerifyVideo KYC and V-CIPPayee change or payment release on a video call
Public footage of an executive becomes a live video persona. The employee sees and hears the CFO, so the instruction looks authorised. A second approver who trusts the same call adds no independent check.
NeuroVerifyNeuroDetectHigh-value payment approvalAccount recovery and the contact centre
A voice cloned from the customer’s public audio, plus answers from breached data, gets a password reset or a new device bound. The FBI reports criminals reaching bank accounts with AI-generated audio of the holder.
NeuroCallBetaNeuroVerifyAccount recoveryPeriodic KYC refresh and device rebinding
After a takeover, the attacker answers the bank’s re-verification with a replayed or synthetic selfie video. The refresh then confirms the attacker as the customer.
NeuroVerifyKYC refreshInsurance onboarding and claims
A synthetic applicant passes the policy video KYC. Later, generated damage photos and an edited invoice support a claim that no adjuster can inspect in person.
NeuroVerifyNeuroDetectClaims evidenceFraud investigation and SAR evidence
FinCEN notes banks often find deepfake identities by re-reviewing account-opening files. That review needs findings an investigator can defend, not a reviewer’s impression.
NeuroDetectMedia investigation
Why your controls pass it
Each control checks that a person is there. None checks that the person is real.
These controls work as designed. A deepfake is built to satisfy exactly the question each one asks.
Face match (1:1)
Does the selfie match the ID photo?
A face swap built from that same ID photo. It matches by design.
Liveness
Is a person moving in front of the camera?
A real-time swap that blinks and turns on cue. An injected stream never meets a camera at all; MAS describes virtual-camera injection by name.
Video KYC agent
Does the person look and answer like the customer?
Answers taken from breached data, given by a face that looks right on a compressed video feed.
Voice biometrics and security questions
Does the caller sound like the enrolled voice and know the answers?
A clone trained on the customer’s public audio. FS-ISAC warns voice recognition must now tell real voices from synthetic replicas.
Executive instruction on a video call
Did someone with authority ask for this payment?
An authority figure who is a rendered face and a cloned voice. In Hong Kong, everyone on the call was fake.
NeuroVerify
Is this face, video or voice real?
One risk score from six signals, including Face Manipulation Detection and AI Generative Model Analysis, before the account opens or the release goes through.
NeuroDetect
What exactly was altered in this file?
Findings down to regions, frames and audio segments, with a forensic report for the investigation or SAR file.
For the product rows, the last column is what the check adds.
One video KYC session
Three checks pass. The account still should not open.
The stream comes from a virtual camera, so liveness sees the motion it expects. NeuroVerify looks at the face itself and finds the swap. The application goes to a person instead of an auto-approve rule.
- Face matchPass
The swapped face was built from the ID photo, so it matches.
- LivenessPass
The swap blinks and turns its head on every prompt.
- Agent questionsPass
Date of birth and address come from breached data.
- NeuroVerify · Face Manipulation DetectionFlag
Blending along the jaw and hairline does not match the skin.
- NeuroVerify · AI Generative Model AnalysisFlag
Generator traces in the face region, frame after frame.
Regulation and guidance
From Washington to Hong Kong, regulators describe the same attack.
What each document says, in plain terms. None of them names a vendor, and using a product does not by itself meet any of them.
- United States13 Nov 2024
FinCEN · FIN-2024-Alert004
Reports a rise since 2023 in suspicious-activity reports describing deepfakes. Its red flags include a third-party webcam plugin during live verification and a photo or video flagged by deepfake detection software. SARs should carry the term FIN-2024-DEEPFAKEFRAUD.
- New York16 Oct 2024
NYDFS · Industry letter on AI cybersecurity risks
Calls AI-enabled social engineering one of the most significant threats to the sector. Advises liveness or texture analysis for biometrics and authentication that deepfakes cannot easily imitate. It adds no requirements beyond Part 500.
- Singapore18 Sep 2025
MAS · Cyber Risks Associated with Deepfakes
Names three impact areas: defeating biometric authentication, social engineering, and misinformation. Recommends testing liveness against deepfake samples, detecting injection, and confirming video-call requests through a different channel.
- Hong Kong3 Sep 2026
HKMA · Remote on-boarding of individual customers
States that deepfake technology has raised the potential for impersonation, online fraud and mule-account networks at scale. Asks banks and SVF licensees to review their remote on-boarding solutions against current fraud intelligence.
- IndiaAmended 9 Jan 2020
RBI · V-CIP in the Master Direction on KYC
Requires a liveness check against spoofing in video-based customer identification, and varied questions so the session is shown to be real-time, not pre-recorded.
- Global industryOct 2024
FS-ISAC · Deepfakes in the Financial Sector
Lists the controls for customer fraud as MFA, deepfake detection, liveness and callbacks to confirm transactions. For executive impersonation it adds authenticated access to internal video conferencing.
Integration
One call at onboarding. The same call at every step-up.
Add it to onboarding
NeuroVerify runs on the selfie or video your KYC flow already captures, through the mobile SDK or REST.
Call it again at step-up
Payee change, limit increase, device rebinding, recovery, KYC refresh. Same call, different policy.
Route on your policy
Continue, recapture, hold for a callback to the number on file, or send to senior review.
Give investigators the file
Flagged sessions, recordings and documents go to NeuroDetect for a forensic report.
SDK-first integration
One API.
Every stack.
NeuroVerify inside your banking app and KYC backend: native SDKs for Android, iOS and Flutter, plus Python, TypeScript, Go, Java and REST. Two lines to integrate; live in under 24 hours.
Sign in for API access$ npm install @neuraldefend/sdk
import { NeuroVerifyClient } from "@neuraldefend/sdk"
const result = await new NeuroVerifyClient().detectImage("selfie.jpg")Mobile SDKs
Native capture apps- KAndroid (Kotlin)
- SiOS (Swift)
- FFlutter
Backend SDKs
Server-side checks- PyPython
- TSTypeScript / Node.js
- GoGo
- JJava
REST API
For everything elseAlready have capture? Send the image to POST /detect/image and get every signal back in one JSON response.
AI agents
MCP serverGive approved AI assistants bounded image and video authenticity tools, with API keys kept outside prompts.
Two lines to integrate<200 ms p95Live in under 24 hours
NeuroDetect runs as a web app, in your private cloud or VPC, as a secure container, on-premises or air-gapped, so case files can stay inside the bank. ISO/IEC 27001 · 27701 · 42001 · SOC 2 Type II. Deployment options
Knowledge centre
Read deeper on deepfakes in financial services.
- Comparison
Top 5 deepfake detection APIs
How the main APIs compare on accuracy, modalities and integration, for teams writing a shortlist.
Read article - Onboarding
Remote bank customer onboarding review
Where a deepfake check sits between document capture, face match and the account decision.
Read article - Mechanism
Liveness vs deepfake detection for KYC
Why a live, moving face and a real face are different questions, and which control answers each.
Read article - Payments
High-value payment approval
Designing a release step that a convincing executive on a video call cannot talk past.
Read article - Contact centre
Financial contact-centre identity review
What an agent should do when the caller sounds right, knows the answers and still is not the customer.
Read article - Mule accounts
Mule account enrolment with synthetic media
How synthetic faces feed the funnel accounts FinCEN describes, and what gives them away later.
Read article - Ongoing due diligence
Periodic KYC refresh and risk reassessment
Re-verification after onboarding is a second door. How to keep it from confirming an attacker.
Read article - Insurance
Insurance customer onboarding review
Applying the same identity checks to policy onboarding before a claim is ever filed.
Read article
Sources
Read the evidence yourself.
Regulators, industry bodies, research and reporting. None of them evaluates or endorses Neural Defend.
- FinCENFIN-2024-Alert004: Fraud schemes involving deepfake media
13 Nov 2024. Rise in deepfake-related SARs since 2023; GenAI identity documents used to open funnel accounts; nine red flags.
- NYDFSCybersecurity risks arising from artificial intelligence
16 Oct 2024. AI-enabled social engineering as a top threat; liveness for biometrics; no new requirements beyond Part 500.
- Monetary Authority of SingaporeCyber Risks Associated with Deepfakes
18 Sep 2025. Biometric bypass, injection attacks, impersonation scams and mitigations, with incident examples.
- Hong Kong Monetary AuthorityRemote on-boarding of individual customers
3 Sep 2026. Supersedes the 2019 and 2021 circulars; names deepfakes as a driver of impersonation and mule networks.
- Reserve Bank of IndiaAmendment to Master Direction on KYC (V-CIP)
9 Jan 2020. Liveness checks against spoofing; varied questions to rule out pre-recorded video.
- FS-ISACDeepfakes in the Financial Sector
Oct 2024. Threat taxonomy and controls for customer fraud, executive impersonation and detection pipelines.
- EBA and ECB2025 Report on Payment Fraud
Dec 2025. €4.2bn EEA payment fraud in 2024; payer manipulation over half of credit-transfer fraud value. Not a deepfake estimate.
- UK FinanceAnnual Fraud Report 2026
Jun 2026. £1.28bn stolen through payment fraud in 2025; £576.4m APP losses, £354.3m returned to victims. Not attributed to deepfakes.
- Deloitte Center for Financial ServicesGenerative AI is expected to magnify the risk of deepfakes
29 May 2024. Projects US gen-AI-enabled fraud losses of US$40bn by 2027, from US$12.3bn in 2023.
- Entrust2026 Identity Fraud Report
18 Nov 2025. Deepfakes in one in five biometric fraud attempts; injection attacks up 40%. Vendor data from its own verifications.
- CNNArup revealed as US$25 million deepfake scam victim
16 May 2024. HK$200m in 15 transfers after a video call with a deepfaked CFO; Arup confirms fake voices and images.
- FBI IC3Criminals use generative AI to facilitate financial fraud
3 Dec 2024. AI-generated audio of account holders used to access bank accounts.
Questions
What banks and insurers ask first.
Compare the top 5 deepfake detection APIsDeepfake detection API for banksDoes NeuroVerify check ID documents?
No. NeuroVerify analyses face images, video and audio. Document authentication stays with your IDV provider. When an investigator needs to examine a document, it goes to NeuroDetect.
Our video KYC already has liveness. What does this add?
Liveness asks whether a person is present. NeuroVerify also asks whether the face is manipulated or generated, using six signals including Face Manipulation Detection and AI Generative Model Analysis. A real-time face swap passes the first question and fails the second. Liveness vs deepfake detection
Can it stop a payment approved on a deepfaked video call?
The control that stops it is procedural: a callback on a known number and separation of duties, as MAS and FS-ISAC both recommend. NeuroVerify adds evidence at the step-up, on a short face or voice sample captured in your app before a payee change or release. NeuroDetect examines a recording afterwards.
What about calls into the contact centre?
NeuroCall, in Beta, analyses the caller’s voice during a live call. NeuroVerify checks a recorded voice sample, for example one captured during a recovery flow. About the NeuroCall Beta
Will a score turn away genuine customers?
Only if your policy says so. Treat a high score as a reason to recapture, call back or review, not to refuse. FS-ISAC notes that customers using speech-assistance tools can look synthetic to a detector, so they need an alternative route. Eligible image requests return in under 200 ms at p95 under defined test conditions.
Does using it make us compliant with FinCEN, MAS, HKMA or RBI?
No product does that. NeuroVerify and NeuroDetect supply evidence and an audit trail; your compliance team decides how they map to each regulator’s expectations. Neural Defend holds ISO/IEC 27001, 27701 and 42001 certifications and a SOC 2 Type II report.
See it on your own sessions
Find the synthetic customers you have already onboarded.
Bring a sample of past video KYC sessions, recovery recordings and claim files, including cases you know were fraud. See what NeuroVerify and NeuroDetect find, and what the evidence looks like on file.