Voice‑clone and deepfake fraud defence
Also called deepfake detection, CEO fraud prevention, synthetic voice detection, payment call-back controls
Cloned voices and deepfake video calls fool staff. Detection software is unproven for small firms; call-back checks are what authorities advise.
People cannot reliably spot a cloned voice: in a 2025 study, clones built from under five minutes of speech passed as human 58% of the time. Detection software is not a fix either; open-source detectors' detection scores (AUC) fell by roughly half on real 2024 deepfakes. In a case reported in 2024, an Arup employee made 15 transfers worth HK$200 million after a faked video call. The FBI's advice is procedural: hang up and call back on a number you find yourself.
At a glance
| Status | In pilots |
|---|---|
| What it does | Combines payment controls (call-backs, code words, two-person approval) with software that scores calls, voice notes or video for signs of synthetic media. |
| Typical cost | Varies by vendor for detection software; the controls law enforcement recommends cost staff time, not licences |
| Data it needs | A verified supplier and staff contact list held outside email, your payment approval rules and, for detection tools, access to call or meeting audio and video. |
| Setup effort | Low for call-back and code-word controls; High for detection software, which has to sit inside phone and video systems. |
| Evidence | Thin. Two peer-reviewed listening studies, one academic detector benchmark, documented cases and law-enforcement alerts; no independent study of detection tools in small firms. |
What is voice-clone and deepfake fraud defence?
Voice-clone fraud uses AI-generated speech to impersonate someone the victim trusts: the owner asking for an urgent transfer, a supplier confirming new bank details, a bank official. Deepfake video takes the same trick onto a video call. The FBI warns that criminals use cloned audio, and real-time video of supposed executives, to get money moved. In South Africa, SABRIC has warned that criminals clone the voices of bank officials. The UK's National Cyber Security Centre assesses that AI will almost certainly make social engineering more effective and harder to detect.
Defence comes in two layers. The first is process: rules under which a voice or a face alone never authorises a payment or a change of bank details. The second is software that analyses audio or video for signs of synthesis. For a small business, the first layer carries almost all of the proven value.
How does it work?
Process controls work by moving verification to a channel the fraudster does not control. The FBI advises:
- Hang up and call back on a number you find independently.
- Agree a secret word or phrase. The FBI suggests this for families; it works as well between an owner and the bookkeeper.
- Confirm through a trusted channel whenever a known contact switches to a new number or platform.
Two-person approval for new payees and bank-detail changes adds a second check that one convincing call cannot clear.
Detection software scores audio, images or video for artefacts of synthesis and flags suspect items for review. FinCEN notes that financial institutions may use it, and separately recommends multifactor authentication and live verification checks, after reporting a rise since 2023 in suspicious activity reports describing suspected deepfake media.
Neither layer helps if staff bypass it under pressure, which is exactly what a convincing impersonation of the boss is designed to produce.
What it needs to work
Defence depends on verification details the fraudster cannot reach and payment rules nobody can override by phone. Detection software is optional; the contact register and approval rules are not.
- A verified contact register for suppliers, banks and senior staff, with phone numbers captured at onboarding and stored in the accounting system's supplier record rather than taken from emails
- Written payment rules: any new payee or bank-detail change needs a call-back on the registered number and a second approver
- Approval limits set in online banking so no single person can release a large or first-time payment
- An agreed code word or phrase for urgent requests from the owner or finance lead
- A short briefing at least once a year for everyone who can move money, using real cases such as Arup
- For detection tools only: access to call, meeting or voice-note audio, and a vendor contract covering how recordings are stored and secured
Most use cases in this guide succeed or fail here, on the data underneath them, before any AI is involved.
What does the evidence show?
The evidence on the threat is strong; the evidence for AI detection is thin. In a 2023 UCL study of 529 people, listeners identified deepfake speech only 73% of the time, and training helped only slightly. A 2025 study found voices cloned from a few minutes of recordings with off-the-shelf software were judged human 58% of the time, against 62% for real voices.
Machines do better, but not well enough. An academic benchmark of deepfakes circulating in 2024 found open-source detectors' accuracy scores fell by about 50% for video and 48% for audio compared with older test sets; commercial tools did better but still trailed human forensic analysts.
Action Fraud, the UK's national fraud reporting centre, gives matching advice on mandate fraud: check any change of bank details using contact details you already hold, not those in the request. We found no independent study measuring whether detection software prevents losses in small businesses.
Detection software vs call-back and approval controls
| Compared on | Deepfake detection software | Call-back and approval controls |
|---|---|---|
| What it catches | Some synthetic audio, images and video, scored by likelihood | Any impersonation, real or synthetic, because it checks the request rather than the voice |
| Error profile | Accuracy falls on new, real-world fakes | Fails when staff skip it under pressure or ring the fraudster's number |
| Cost model | Licence; varies by vendor | Staff time |
| Best for | Banks and firms that verify many inbound callers or customers | Every business that pays suppliers or staff by electronic transfer |
Status and evidence
Rated pilots: detection tools exist and FinCEN notes financial institutions may use them, but an independent benchmark shows accuracy falling sharply on real-world fakes, and we found no evidence of results in small firms. Evidence for the AI layer is thin: one benchmark and two listening studies. The case for call-back controls rests on law-enforcement guidance rather than measured trials.
Where does it go wrong?
- The call-back goes to the fraudster. Staff who ring the number in the email or on the new invoice reach the criminal. Call back only on contact details you already hold, such as those captured when the supplier was set up.
- Doubts get talked away. The Arup employee was suspicious of the first email, then set the doubts aside after a video call with people who looked and sounded like colleagues.
- Voice notes treated as authority. A voice message in the owner's voice proves nothing; the FBI warns that AI clones can sound nearly identical to real contacts and advises verifying requests independently.
- Trusting the detector. Detection accuracy falls sharply on real-world fakes, so a clean score is not permission to pay.
- Public voice samples. Owners who post videos and podcasts hand criminals material; convincing clones needed only a few minutes of speech.
- Controls that only cover payments. The same tricks are used to get logins, so never give a one-time code to anyone who asks for it.
Numbers it should move
Ask the demo business
Northgate Supply is a fictional wholesale distributor of catering and packaging supplies: 28 staff, R46.6m in sales over the last year. Its data is generated for this demo, to 30 September 2026.
2 pairs of payments go to the same supplier for the same amount within two weeks, worth R216 436 in total. The references differ only in punctuation, which is how duplicates usually slip past a manual check.
How this was worked out
Compares every pair of supplier payments made within 14 days of each other, flagging the same supplier and the same amount to the cent.
Where the data usually lives
Where the records this use case needs are usually kept in a small business. It is not a list of recommended tools.
Frequently asked questions
Can someone really clone my voice from a few videos online?
Yes. In a 2025 study, voices cloned with off-the-shelf software from under five minutes of speech were judged human 58% of the time, close to the 62% for real voices. The FBI warns that AI voice clones can sound nearly identical to real contacts and advises confirming any request through separate contact details.
What is the cheapest way to stop deepfake payment fraud?
A rule that no payment or bank-detail change is approved on the strength of a call, voice note or video alone. Verify by hanging up and calling back on a number you find independently, not one in the request. The FBI suggests families agree a secret word; an owner and bookkeeper can do the same for urgent requests. These cost staff time, not software. Detection tool prices vary by vendor, and they add a layer, not a substitute.
Do I need an IT team to defend against voice-clone fraud?
Not for the controls that matter most. Call-backs, code words and two-person approval are written procedures. Multifactor authentication on email and banking, which the FBI also recommends, is a settings change an outsourced IT provider can make. Detection software that sits inside phone or video systems does need technical setup and a vendor contract.
Are deepfake scams happening in South Africa?
Yes. SABRIC, the South African Banking Risk Information Centre, has warned that criminals use AI to clone the voices of bank officials and to copy bank communications closely. We found no published South African statistics on business losses to deepfakes specifically. In the UK, the NCSC assesses that AI will almost certainly make social engineering more effective and harder to detect.
Does recording calls to check for deepfakes raise POPIA or UK GDPR issues?
Yes. A recording of someone's voice is personal information. If a vendor analyses recordings for you, POPIA requires a written contract obliging it to keep the information secure, and the Act requires it to tell you immediately if it suspects unauthorised access. UK GDPR sets similar processor contract terms, which the ICO explains. Decide how long recordings are kept and who can hear them. This is not legal advice.
More in risk, compliance & security
Sources
- 1Engineering giant Arup falls victim to deepfake scamData Protection Office, The University of Hong Kong, 2024News
- 2Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud (I-120324-PSA)FBI Internet Crime Complaint Center (IC3), 2024Government
- 3Senior US Officials Impersonated in Malicious Messaging Campaign (I-051525-PSA)FBI Internet Crime Complaint Center (IC3), 2025Government
- 4Mandate fraudAction Fraud (National Fraud & Cyber Crime Reporting Centre, UK), 2026Government
- 5FinCEN Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions (FIN-2024-Alert004)Financial Crimes Enforcement Network, US Treasury, 2024Government
- 6Humans unable to detect over a quarter of deepfake speech samplesUniversity College London, 2023Study
- 7Humans flunk the Turing test for voices as bots get chattierThe Register, 2025News
- 8Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024arXiv (Chandra et al.), 2025Study
- 9AI-powered scams drain South Africans' bank accounts — here's how to protect your moneyDaily Maverick, 2025News
- 10The near-term impact of AI on the cyber threatNational Cyber Security Centre (UK), 2024Government
- 11Protection of Personal Information Act 4 of 2013, section 21: Security measures regarding information processed by operatorActs Online (South African statute), 2013Government
- 12What needs to be included in the contract?Information Commissioner's Office (ICO), 2026Government
This entry is educational. It is not legal, financial or professional advice, and it does not recommend any vendor. Figures attributed to vendors are their own claims.
- 1Engineering giant Arup falls victim to deepfake scamData Protection Office, The University of Hong Kong, 2024News
- 2Criminals Use Generative Artificial Intelligence to Facilitate Financial Fraud (I-120324-PSA)FBI Internet Crime Complaint Center (IC3), 2024Government
- 3Senior US Officials Impersonated in Malicious Messaging Campaign (I-051525-PSA)FBI Internet Crime Complaint Center (IC3), 2025Government
- 4Mandate fraudAction Fraud (National Fraud & Cyber Crime Reporting Centre, UK), 2026Government
- 5FinCEN Alert on Fraud Schemes Involving Deepfake Media Targeting Financial Institutions (FIN-2024-Alert004)Financial Crimes Enforcement Network, US Treasury, 2024Government
- 6Humans unable to detect over a quarter of deepfake speech samplesUniversity College London, 2023Study
- 7Humans flunk the Turing test for voices as bots get chattierThe Register, 2025News
- 8Deepfake-Eval-2024: A Multi-Modal In-the-Wild Benchmark of Deepfakes Circulated in 2024arXiv (Chandra et al.), 2025Study
- 9AI-powered scams drain South Africans' bank accounts — here's how to protect your moneyDaily Maverick, 2025News
- 10The near-term impact of AI on the cyber threatNational Cyber Security Centre (UK), 2024Government
- 11Protection of Personal Information Act 4 of 2013, section 21: Security measures regarding information processed by operatorActs Online (South African statute), 2013Government
- 12What needs to be included in the contract?Information Commissioner's Office (ICO), 2026Government