zlacker

>>leonry+(OP)
Best of luck to the author! My understanding is that anything that makes large file sharing easy and anonymous rapidly gets flooded with CSAM and ends up shuttering themselves for the good of all. Would love to see a non-invasive yet effective way to prevent such an incursion.

>>Arubis+3s
If governments and big tech want to help, they should upload one of their CSAM detection models to Hugging Face, so system administrators can just block it. Ideally I should be able to run a command `iscsam 123.jpg` and it prints a number like 0.9 to indicate 90% confidence that it is. No one else but them can do it, since there's obviously no legal way to train such a model. Even though we know that governments have already done it. If they won't give service operators the tools to keep abuse off their communications systems, then operators shouldn't be held accountable for what people do with them.

>>jart+cA1
This would potentially let somebody create a "reverse" model, so I don't think that's a good idea.

Imagine an image generation model whose loss function is essentially "make this other model classify your image as CSAM."

I'm not entirely convinced whether it would create actual CSAM instead of adversarial examples, but we've seen other models of various kinds "reversed" in a similar vein, so I think there's quite a bit of risk there.

>>miki12+PN1
Are you saying someone will use it to create a CSAM generator? It'd be like turning smoke detectors into a nuclear bomb. If someone that smart wants this, then there are easier ways for them to do it. Analyzing the detector could let you tune normal images in an adversarial way that'll cause them to be detected as CSAM by a specific release of a specific model. So long as you're not using the model to automate swatting, that's not going to amount to much more than a DEFCON talk about annoying people.

>>jart+7Q1
I think the point is generating an image that looks normal but causes the model to false positive and the unsuspecting person then gets reported

>>throwa+x02
If you have a csam detection model that can run locally, the vast majority of sysadmins who use it will just delete the content and ban whoever posted it. Why would they report someone to the police? If you're running a file sharing service, you probably don't even know the identities of your users. You could try looking up the user IP on WHOIS and emailing the abuse contact, but chances are no one is listening and no one will care. What's important is that (1) it'll be harder to distribute this material, (2) service operators who are just trying to build and innovate will be able to easily protect themselves with minimal cost.

>>jart+C03
Someone send you a meme that looks like a meme, you share it through messenger, the meme looks like something else to the messenger, messenger reports you to NCMEC. It's NOT police but they can forward it to police. As a side effect NCMEC gets overloaded helping more of real abuse continue.