What Is a Deepfake?
A deepfake is video, audio, or imagery that machine learning has created or manipulated (opens in new tab) to depict a real person doing or saying something they never did. The word combines "deep" from deep learning with "fake."
The machine-learning step separates a deepfake from a cheapfake, a conventional edit such as slowing a speech or showing footage out of context.
How deepfakes work
A deepfake model is trained on reference content (opens in new tab), such as photos, video, or recordings of the target, then reproduces that person in situations that never occurred. Three model families do most of the work. Autoencoders compress a face into a compact code, and a classic face swap (opens in new tab) uses a shared encoder with one decoder per identity so one person's expressions can be decoded as the other person's face.
Generative adversarial networks pit a generator against a detector over many cycles until the fake passes. Diffusion models learn to reverse noise (opens in new tab) that is added to real images step by step, which lets them generate new images from noise.
Voice cloning follows the same logic. Neural codec language models treat speech as a sequence of audio tokens, and one such model, VALL-E, synthesizes personalized speech (opens in new tab) from a three-second recording. Real-time tools carry this onto live calls: a Unit 42 researcher (opens in new tab) with no prior experience built a working real-time deepfake in just over an hour on cheap consumer hardware.
Picture the result at executive scale. An attacker clones a CFO's voice from earnings-call audio, drives a face swap from conference footage, and joins a scheduled call to push through a transfer.
Why deepfakes are hard to stop
A deepfake targets the checks people fall back on when other controls have passed, such as a familiar face on a video call or a known voice on the phone. It also widens the attack surface: paired with audio and video deepfakes, any conversation over an unauthenticated or poorly vetted channel becomes an easy target, email included.
Detection is not a safe fallback. AI tools built to detect synthetic content (opens in new tab) can be ineffective and unreliable, so defense needs other technical measures too. Detectors trained on known generators tend to degrade on new ones, and provenance standards such as C2PA Content Credentials only help when content carries them.
An attacker can strip the metadata, render a watermark unreadable, and register a new manifest, a scenario the C2PA specification (opens in new tab) itself describes.
Deepfake tooling also sells as a service. Criminals have shifted from custom underground services to mainstream creation platforms (opens in new tab) that offer real-time streaming manipulation and multilingual voice cloning cheaply, and turnkey KYC-bypass kits (opens in new tab) such as ProKYC, sold through Telegram channels, defeat document-plus-selfie onboarding.
Deepfakes now appear in CEO fraud (opens in new tab), where fraudsters impersonate the CEO to trick employees. Campaigns also span channels: since 2023, text and AI-generated voice messages impersonating senior U.S. officials have built rapport before moving targets to another messaging platform (opens in new tab).
Types of deepfakes
- Pre-recorded video. Creators produce face swaps or reenactments in advance and publish them as ads, clips, or endorsements. In November 2025, Berkshire Hathaway warned (opens in new tab) that YouTube videos used AI-created images and an impersonated voice to present alleged comments by Warren Buffett.
- Voice clones. Synthetic audio that mimics a known speaker and drives vishing calls, voicemails, and the senior-official campaigns described above. See AI voice cloning for the mechanism.
- Real-time deepfakes on live calls. Tools render face, voice, or both during a Zoom or Microsoft Teams meeting, as in the Singapore case below. Attackers use them for executive impersonation and for fraudulent job candidates (opens in new tab), and North Korean IT workers have been observed using face-swapping in video interviews (opens in new tab).
- Image deepfakes, morphs, and forged identity documents. Attackers submit synthetic faces or altered documents to remote identity checks. Morphed photos (opens in new tab) can deceive face recognition systems, and in a case Hong Kong authorities announced in June 2026 (opens in new tab), a syndicate replaced the portraits on lost identity cards with deepfakes to pass online banking facial-recognition checks.
How to defend against deepfakes
Out-of-band verification defeats a deepfake that has already passed the eye and ear test. For transfers and vendor-payment changes, confirm by phone using a known telephone number (opens in new tab), never the number or link in the request, and write that step into company policy before a request arrives. Use phishing-resistant MFA and live verification checks on identity verification, and limit who can initiate credential resets at the help desk.
Reduce the raw material. Limit the public images and voice content attackers can train on, starting with executives, and sign official executive media with Content Credentials (opens in new tab) so a fake is harder to pass off as legitimate. In hiring, ask video candidates to wave a hand (opens in new tab) in front of their face, which can prompt a malfunction in AI-generated video.
Run tabletop exercises that include the executives most likely to be impersonated, update incident response playbooks for multimodal impersonation (opens in new tab), and correlate impersonation reports with account recovery events, new devices, privilege changes, and financial transactions. When a deepfake campaign reaches customers or the public, removing the impersonating content is a core brand-impersonation control.
How Doppel helps
Doppel is the Agentic Social Engineering Defense (SED) platform unifying Digital Risk Protection and Human Risk Management. It treats a deepfake as one artifact of a wider impersonation campaign, alongside the spoofed profile, the lookalike domain, and the cloned voice behind the call.
Doppel Executive Protection detects impersonation profiles and deepfake content tied to leadership, then dismantles the attacker infrastructure behind them.
The Doppel Threat Graph correlates deepfake content, spoofed profiles, and cloned-voice calls into a single campaign view, and agentic AI prioritizes and executes takedowns across registrars, hosts, social platforms, and ad networks, so analysts handle only the complex escalations that need human judgment.
The same threat data trains the people attackers target. Simulation runs live voice simulations with deepfake or AI-generated voices, including inside Microsoft Teams meetings, and Security Awareness Training can feature deepfake videos of your own executives. Every detected threat converts to an employee simulation in one click, and every dismantled campaign strengthens detection for every customer, which is how an organization becomes too costly to attack.
Request a demo to see how the platform maps and dismantles a live deepfake campaign and converts it into an employee simulation.
Frequently asked questions about deepfakes
What is a deepfake?
A deepfake is a video, image, or audio recording that machine learning has generated or altered to show a real person doing or saying something they never did. Machine or deep learning is the defining ingredient, which is why the EU AI Act (opens in new tab) defines a deep fake as AI-generated or manipulated content that resembles real people, places, or events and would falsely appear authentic. Ordinary edits, such as slowing a clip, create cheapfakes instead. Creating synthetic content is not inherently illegal (opens in new tab), but attackers use it to commit fraud and extortion.
What is a deepfake in cybersecurity?
In cybersecurity, a deepfake is a social engineering tool: attackers clone a trusted person's voice or face, often an executive's, and use it to authorize fraudulent transactions (opens in new tab), extract credentials, or pass identity checks. Common forms include cloned-voice calls to employees, deepfaked executives on video calls, synthetic faces that defeat know-your-customer verification, and fake job candidates in remote interviews. Defense relies on out-of-band verification and phishing-resistant authentication, because the fake is often too good to spot by eye.
What is the difference between a deepfake and a cheapfake?
Machine learning is the dividing line. A deepfake is created or manipulated with machine or deep learning, while a cheapfake uses techniques that do not involve machine learning (opens in new tab), such as speeding or slowing a clip or splicing footage out of context. Cheapfakes take less technical skill, but the machine-learning step is what lets a deepfake put convincing new words in a specific person's mouth. Both sit inside the broader category of manipulated media.
What is an example of a deepfake attack on a business?
In March 2025, a finance director at a multinational company in Singapore was contacted on WhatsApp (opens in new tab) by a scammer impersonating the CFO, then joined a Zoom call the next day where deepfaked company officials, including the CEO, discussed restructuring the firm's regional business. The director transferred over US$499,000 before Hong Kong's Anti-Deception Coordination Centre, working with Singapore's Anti-Scam Centre, withheld the full amount on March 28. The sequence is the pattern to watch for: contact on a messaging app, a pivot to a video call, and an urgent, confidential financial request. Verifying the request through a known phone number breaks it.