żěè¶ĚĘÓƵ

AI-generated deepfake faces could help protect privacy on social media

Algorithms that generate fake faces could be used to change people’s appearance in photos they don’t want to be shared
Multiple versions of a group photo with faces altered
The AI tool can alter each face in a group photo depending on who is viewing it
Umur Ciftci / Binghamton University

Social media users could use a face-generating artificial intelligence system to hide their identity in other people’s photos.

Platforms such as Facebook and Instagram let users decide whether they are tagged in photos. However, there is no way to stop other users sharing those photos, which is a problem – especially given the increasing power of facial recognition AI technology, which can search the web for photos of any individual.

“Even if you untag your face, [the pictures] are online forever,” says at Intel Labs. “Faces are the most important biodata, and we need to have control over it.”

Demir and her colleagues have developed a method that uses deepfakes – fake faces generated by artificial intelligence – to obscure a person’s appearance on social media depending on who is viewing the photo.

The system, called My Face My Choice, stores a digital representation of a person’s face and produces an AI-generated synthetic version according to that person’s privacy settings. The synthetic face can be modified to make it more or less similar to a person’s real face.

The tool is intended to give people anonymity by hiding their faces from anyone they don’t want to see them. But it could also have a much bigger benefit for online privacy, by undermining facial recognition programs that harvest photos of faces from the internet to train such systems. These programs use a stored catalogue of unique face data, known as a “face space”, to match photos with people they have seen before.

“Instead of trying to decide between 8 billion people, let’s say all of us have thousands of other faces, then it actually explodes the face space of the facial recognition algorithms,” says Demir.

Using deepfakes as a privacy enhancing tool is a significant idea, says at University College London. “I think the results in the paper show that it’s achievable,” he says.

However, there are still technical hurdles to be overcome before it can be deployed on a very large social network like Facebook, such as storage and security needs. Also, it is unclear whether there would be enough demand from social media users who want their face obscured to strangers, says Griffin.

Reference:

Topics: AI / Privacy / Social media