Deepfake
Deepfake
What is a deepfake?
- Faces and voices synthesized to look real
- Used in creative work, but a fraud and disinformation threat
- Built on GANs and diffusion models
What deepfakes have really changed is not video editing technology but the status of evidence. A video showing a face, or a call with the person’s own voice, used to substitute for a verification step by itself. Now work rules have to be rewritten on the premise that a face and a voice no longer prove identity. The rise in wire requests impersonating executives is not only because the technology got more sophisticated; the gap is that organizations still treat “I heard the voice, so it’s fine” as grounds for approval.
There are three main ways to make one: swap types that replace only the face in existing footage, lip-sync types that move the mouth and expressions to match speech, and fully synthetic types that generate a person from scratch. Voice cloning has the lowest barrier to entry. A few dozen seconds of publicly available audio is enough to copy speech habits and intonation. This is also where the line is drawn against CGI, retouching, and AI-generated images. The moment a real person’s identity is borrowed, it stops being creative work and becomes impersonation. That is why the response is shifting from a single detection model toward content provenance standards (such as C2PA), which embed source information at the capture or generation stage.
The most common mistake in practice is the belief that 'you can tell by looking.' People are taught to watch for blinking or awkward edges around the neck, but those cues rarely hold up against the latest models, and on low-quality video calls they vanish altogether. Feeling safe because 'our executives have never been on TV' is risky too. A recruiting video, a YouTube greeting, or an investor-briefing recording gives attackers more than enough raw material in just a few dozen seconds. Some organizations consider the job done once a detection tool is installed, but detection only lowers the odds. Incidents are prevented only when a rule backs it up: irreversible actions such as wire transfers, contract signing, and HR notices must always be reconfirmed through a separate channel.
John KimAuthor of “B2B Marketing Design” and CEO of Nanum Consulting
Hyerang ParkContent storytelling and AI-driven brand content specialistThe term was coined around 2017 as 'deepfake' (deep learning + fake) when face-swap videos spread online, and it became a social issue as generation technology advanced.
Consider a hypothetical case. Company A, an auto-parts maker with 620 employees and KRW 180 billion (≈ US$134,670,100) in annual revenue, sends funds to three overseas subsidiaries. Last November, a single voice message impersonating the head of its Vietnam subsidiary nearly cost it KRW 42 million (≈ US$31,400). The finance team sensed something was off and stalled for 30 minutes to stop it, yet the wire-approval policy still contained only the phrase 'confirm by phone with the responsible manager.'
- At next week’s meeting, Kim, a senior manager on the finance team, reported, “That voice, down to the intonation and the way the sentences trailed off, was the CEO.” The head of administration ordered a fresh review of the rules.
- The security lead, Manager Lee, pulled 17 clips of executive voices, 42 minutes in total, from the company YouTube channel and recruiting videos, then played them back as-is in front of the leadership team.
- The finance team changed its procedure so that overseas transfers of KRW 20 million (≈ US$15,000) or more can no longer proceed on phone or video instructions alone. It now calls the requester back on a pre-registered personal mobile phone and checks an agreed confirmation phrase.
- The training team built a 30-minute internal course that played synthetic and real voices side by side and asked participants to pick the real one. First-round accuracy was only 38%, and the room was unsettled for a while.
- Six months later, 9 suspected impersonation reports had been filed and 3 were confirmed as real. There were 0 actual wire-fraud losses, and compliance with the callback procedure rose from 61% in the first month to 97%.