How anti-deepfake pixels stop fake nude photos before they're made

According to South China Morning Post, a South Korean university has started embedding graduation photos with technology designed to stop those images being turned into sexually explicit deepfakes. Seoul National University adopted the tool after a scandal two years ago, when portraits of female graduates were used to create manipulated sexual images.

The tool is called StealCut Protect. It was built by a start-up founded by Seoul National University alumni, and this is the first time it has been used on a university graduation album in South Korea. The company gave it to the university free of charge.

Where the deepfake actually starts

Most people imagine a deepfake starting with software. It actually starts with a photo. A deepfake generator needs a clear image of a face to analyse its structure before it can build a manipulated version. The photo is the raw material. Without a usable photo, there is nothing to manipulate.

This is the step StealCut Protect targets. It inserts pixel-level signals into a photo, changes so small they are largely invisible to a person looking at the image. A human viewing the photo sees a normal graduation portrait. A deepfake model reading the same file sees something else. The signals disrupt the facial-structure analysis the generator depends on.

The result, according to the source, is that any attempt to manipulate a protected photo produces a distorted or unidentifiable face. The generator still runs. It just fails to produce a usable fake.

What this shows about the scam

The 2023 scandal at the same university shows where the vulnerability sits. It was not a hacked account or a stolen file. It was a photo that existed publicly, in a form any deepfake tool could read cleanly. The scam does not depend on tricking a person into handing over an image. It depends on the image itself being legible to the software that misuses it. Protecting the file at the point of creation, before it is ever shared, closes that step without requiring the person in the photo to do anything differently afterwards.

That is a different kind of defence to the ones usually described in scam coverage. It is not about a call to make or a link to avoid. It is about what a photo is allowed to be readable by, before it ever leaves the photographer's hands.