What is deepfake fraud?
Definition: Fraud that relies on synthetic voice, video or images, created or altered with artificial intelligence, to pass as a real person and obtain money, data or access.
Deepfake fraud is the use of synthetic audio, video or images, produced or altered by artificial intelligence, to impersonate a real person in order to obtain money, information or access. The word deepfake comes from deep learning, the technique used to build the fakes. Targets include company finance staff, relatives of the person imitated, bank customers, and the identity checks that banks use to open accounts.
It is a branch of AI-enabled fraud and a modern form of social engineering. Earlier impersonation relied on a plausible email or a convincing accent. Generative AI lets an attacker copy a real voice from a short recording and produce a live or pre-recorded video that appears to show a familiar face, which removes many of the cues people used to rely on to judge whether a request is genuine.
How it works
Attackers usually start with public material: conference talks, interviews, social media clips and company videos supply the voice and face samples. Software trained on those samples can then generate speech in the target's voice, swap a face onto a live video feed or build a fully synthetic person. The output is combined with ordinary social engineering, such as urgency, secrecy and a plausible reason for an unusual payment.
The main patterns seen by banks and regulators are:
- Executive impersonation, in which a cloned voice or a fake video call instructs an employee to make an urgent transfer, a more convincing variant of business email compromise and vishing.
- Family emergency calls in which a cloned voice of a relative asks for money, a form of impersonation scam.
- Fake celebrity or official endorsements in videos that promote bogus schemes, feeding investment fraud.
- Attacks on customer onboarding, where a synthetic face or altered video is presented during identity verification to open accounts, often linked to synthetic identity fraud and mule accounts.
Notable cases
In 2024 the engineering firm Arup confirmed that it had been the victim of a deepfake fraud in Hong Kong. An employee was reported to have joined a video call in which the apparent chief financial officer and colleagues were fabricated, and went on to authorise transfers totalling about 200 million Hong Kong dollars, roughly 25 million US dollars. An earlier, widely reported case in 2019 involved a cloned voice of a parent company's chief executive used to request a transfer of about 220,000 euros from a UK energy firm. Neither case required breaching a computer system; both relied on trust and on a payment process that could be completed on a single instruction.
Authorities have issued warnings. In November 2024 the US Financial Crimes Enforcement Network (FinCEN) alerted financial institutions to fraud schemes using deepfake media, and asked them to cite a specific key term in suspicious activity reports on such cases.
Red flags
No single sign proves a fake, but patterns raise suspicion. These include an unusual request to bypass normal approvals, pressure to act at once and keep the matter confidential, a demand to use a new account or channel, a call that cannot be returned on a known number, odd lip movement, flat or repetitive speech, lighting or edge distortions in video and a participant who avoids turning their head or answering unexpected questions. Because detection by eye or ear is unreliable and improving fakes erode these signs, institutions treat them as prompts to verify rather than as proof.
Defences
Organisations rely mainly on process rather than on spotting the fake. Standard controls are call-back verification on a number already held on file, dual authorisation of payments above set limits, agreed code words for urgent requests, strict rules for changing payee details and a rule that no payment is released on the basis of a call or video alone. These sit within wider internal control and fraud detection programmes.
Against synthetic faces during onboarding, providers use liveness detection, which asks for a live action or checks for signs of replay, together with biometric verification, document authenticity checks and device signals. Banks also apply machine learning models to payments, so that an instruction that follows the pattern of a scam can be held. Where a customer is persuaded to send money willingly, the loss falls within authorised push payment fraud, with reimbursement rules that differ by country.
Legal treatment
Deepfake fraud is usually prosecuted under existing laws rather than a dedicated offence: in the United States, the federal wire fraud statute and state laws on identity theft and impersonation, and in other jurisdictions fraud and computer misuse offences. Some countries and regions have added rules on labelling or banning deceptive synthetic media, though their scope and timing vary and are changing.
Ref 1140000221