Reduce deepfake risk before your profile photo becomes someone else’s source material

What a generator needs from your photograph

Synthetic video and image models work better with more reference material, higher resolution, and more angles. A single clean, high-resolution, well-lit frontal headshot is a strong starting point. A dozen photographs of the same face from different angles is considerably better than one.

So the levers that reduce deepfake risk are mostly about supply: how many usable reference images of you exist publicly, how good they are, and how easy they are to collect in one place. A face-search engine is what makes the last part trivial, which is why exposure and deepfake risk are related problems even though they are not the same problem.

The honest limit of cloaking here

Adversarial cloaking of the Fawkes kind was designed to disrupt facial recognition — the task of deciding whether two faces are the same person. Generation is a different task with different models, and a perturbation tuned against recognition embeddings is not a general defence against a generator.

It is not nothing: degrading the quality of the reference material degrades what can be built from it, and making your photographs harder to collect via face search reduces how many references an attacker assembles. But anyone telling you that cloaking a profile picture prevents deepfakes is overselling it, and this page is not going to.

What actually reduces the risk

Reduce the number of high-resolution frontal photographs of you that are publicly reachable. Where a photograph has to be public, there is rarely a reason for it to be published at the largest size you have. Get indexed copies de-listed from face-search engines, because that is what turns scattered images into a collectable set.

And treat the downstream problem separately. Most harm from a synthetic video of an executive comes from someone acting on it quickly. Verification habits inside an organisation — a second channel for anything involving money or credentials — do more than anything you can do to an image file.

Where protection does fit

On the photograph you are about to publish, as one of several things that make you a marginally harder target, alongside removal and alongside publishing at a sensible resolution. As the whole of a deepfake strategy, it does not fit at all.