Hacker News Reader

Comment by Pxtl | Hacker News Reader

Intresting and related, a team made a neat "face depixelizer" that takes a pixelated image and uses machine learning to generate a face that should match the pixelated image.

What's hilarious is that it makes faces that look nothing like the original high-resolution images.

https://twitter.com/Chicken3gg/status/1274314622447820801

View on HN

Interesting... Neat... Hilarious... In light of the submission and the comment you're responding to, these are not the words I would choose.

I think there's genuine cause for concern here, especially if technologies like these are candidates for inclusion in any real law enforcement decision-making.

That should be called a face generator, not a depixelizer.

Basically. The faces look plausible but less useful than the original blurred image.

What's sad is that a tech entrepreneur will definitely add that feature and sell it to law enforcement agencies that believe in CSI magic: https://www.youtube.com/watch?v=Vxq9yj2pVWk

And another entrepreneur can add a feature to generate 10 different faces which match the same pixelation, and sell it to the defence.

A better strategy might be to pixelate a photo of each member of the jury, than de-pixelate it through the same service, and distribute the before and after. Maybe include the judge and prosecutor.

Doubt that many people can afford to hire an expert witness, or hire someone to develop bespoke software for their trial.

Ironically, if the police had used and followed the face depixelizer then we may not have had the false arrest of a black man - not because of accuracy but because it doesn't produce many black faces

I wonder if this is trained on the same, or similar, datasets.

One of the underlying models, PULSE, was trained on CelebAHQ, which is likely what the results are mostly white-looking. StyleGAN, which was trained on the much more diverse (but sparse) FFHQ dataset does come up with a much more diverse set of faces[1]...but PULSE couldn't get them to converge very closely on the pixelated subjects...so they went with CelebA [2].

[1] https://github.com/NVlabs/stylegan [2] https://arxiv.org/pdf/2003.03808.pdf (ctrl+f ffhq)

Reply on news.ycombinator.com