Show HN: An AI model that busts AI faces

[albertbui]

Hey HN, I'm an AI enthusiast and I am launching apps that use ML to solve problems that we all have.

I realized that there are a lot of deepfake faces on youtube and social media, so I thought it would be useful (and fun!) to have a tool that can bust those AI faces.

So I built DeeFace for 2 reasons: 1. Check if the face you're looking at is real or not 2. For fun! It's something that I from half a year ago would look at and aspire to build someday. I think it's soemthing anyone who's just starting out with ML would appreciate.

I hope this tool is fun for you to use as it was for me to build. Would love to hear your feedback on this.

Albee

[drivingmenuts]

Turn this into an AI Blocker - create a browser add-on that detects AI faces and reports it to a central site. If a channel gets enough face-detected signals, then put them on a blocklist that users can subscribe to.

No idea how one would get off the list if there was a false positive or discontinuation of AI usage. Digital proof is basically useless at this point.

Personally, I'd just leave them on there forever, since YouTube channels are basically fungible, but some people don't go for that approach.

[albertbui]

This is indeed a very clever application for this tool, automatically detects AI faces and takes actions for the user. Will look more into this. Thanks for the suggestion!

[Remmy]

My trained LoRas seem to defeat it pretty easily, but that's introducing extremely specific facial data. It does well at generic AI faces however.

[albertbui]

Hi Remmy, this is definitely something that must be addressed when this tool becomes popular and people try to bypass it when creating AI faces (like using trained lora). Thanks for pointing that out!