One year ago I developed the first (and from what I know, still only) real-time CSAM detection tool for the fediverse. This has been in use by this instance and recently the real-time version was put in use by lemmy.world. Unfortunately the false-positive rate was a tad too high as this was still using my original implementation in horde-safety. But through our demands in the AI Horde, we've had to constantly tweak and improve it over the past year and thereofre we've had an improved checker for a while, but not used in fedi-safety.
Unfortunately I haven't had the time/motivation to update into it recently so lemmy.world pinged me about its false positive rate being a tad too high, I felt it was a good time to do so.
So now horde-safety has been updated and it should already be more accurate. The admins of lemmy.world already put it into production and they have the most demand, so they'll report back with their findings in a week. If this is not sufficient for lemmy's purpose, I have some other ideas for tweaking it.
And yes, memes and pressure on the admins is what caused me to look into it, but remember we're all just volunteers here. I would have looked into it if y'all had asked nicely as well ;)
Speaking of volunteers, if you want to support my work in providing tooling for lemmy and the Fediverse, feel free to send some support my way which covers all of my FOSS project work.
Curious, how do you evaluate the performance without breaking the law?
Looking at this it looks like the author just... has csam to evaluate with: https://github.com/Haidra-Org/horde-safety/blob/main/tests/test_csam_checker.py
Guess we don't know the laws where they live though. Or where they run the program.
It seems like it could be legally problematic, but I'm not sure what the alternative would be other than accepting the privacy/autonomy nightmare of funneling all traffic through a government affiliated centralized service.
I didn't delve very deep, but it seems it uses a pre trained model that classifies images with anime tags (loli, lewd...) and gives some weights to those. I guess at some point the author will review the results of the real images on Lemmy and use them to tweak those weights, which I understand might be way at least temporary illegal (they could destroy the image and keep only the transformed version, which is "impossible" to turn back into the original image, do it still works as training data)
DeepDanbooru is the model of you're interested.
Deepdanbooru is one of the two models. We also use open ai clip
Well, if that was my only mistake then your code was surprisingly easy to follow (rule number one of github: never read the readme.md)
How do you deal with tweaking? Do you get random samples from Lemmy? Anything you can share about the legal aspect.of it?
horde-safety is there primarily to protect the AI Horde. We have no shortage of creeps and pedos trying to use our crowdsource resources.