Protecting a face against deepfakes

Shield the photo, and the fake falls apart

SHIELD adds a barely visible signal to your photographs. Give a shielded photo to a leading generative AI image tool to fake you, and the output collapses into a distorted image that is clearly not genuine. It does this today on five of the six leading tools we test. Work on the sixth is under way.

Product

SHIELD

Collapses

Five of the six leading generative AI image tools

In progress

The sixth

Visibility

Barely visible, judged by the person it protects

For

Individuals and public figures; R&D partners and funders

One public photo is all a deepfake needs

Today’s image generators can rebuild a person from a single photograph: a new pose, a new setting, a scene that never happened. Every headshot, press photo and profile picture you publish is raw material for someone who wants to fake you.

Detection comes after the damage is done. SHIELD works before it: it protects the photo itself, so the fake never comes out right.

Shield your photo

SHIELD adds a signal to the face in your photograph, tuned to stay barely visible to the person in it.

Publish as usual

Post it, send it, put it in your press kit. The shield is built to survive the compression and resizing photos go through when they are shared.

The fake falls apart

When someone feeds the photo to a generative AI image tool to fake you, the output collapses into a distorted image that is clearly not genuine.

5 of 6

leading generative AI image tools collapse on a shielded photo

1. 01

   Catastrophic collapse, not a weaker likeness.

   On five of the six leading tools we test, the output from a shielded photo is a distorted image that is clearly not genuine.

2. 02

   The sixth is next.

   We are running experiments on the remaining tool now, and will report when the result holds.

3. 03

   Open-source face swap too.

   The shield also holds against the open-source face-swap path that many deepfake apps are built on.

4. 04

   Barely visible.

   The person being protected sets the visibility bar, and judges it blind, on unlabelled images.

Individuals and public figures

Protect the photos you publish

Executives, politicians, performers, journalists and anyone whose face is public. We shield your portraits, headshots and press images before they go out.

R&D partners and funders

Help close the last gap

We are looking for research partners and funders to crack the sixth tool, extend SHIELD to more people and photographs, and validate it independently.

We publish only what survives an audit

1. 01

   Tested on the tools people actually use.

   Shielded photos go to the leading generative AI image tools with an ordinary editing request, the way a deepfake would be made.

2. 02

   Checked by recognisers it was never built against.

   Identity is measured with face-recognition models from more than one family, including one never used to build the shield, so a result cannot be the shield reading itself back.

3. 03

   Judged by the person it protects.

   Visibility is the protected person’s call, made blind on unlabelled images.

4. 04

   Audited before it is claimed.

   Every result is cross-checked against its own data. Anything that does not survive is withdrawn.

SHIELD protects the photos you shield before you publish them. It cannot protect photos already online, or photos other people take and post.