In the past months, there’s a been a issue in various instances where accounts would start uploading blatant CSAM to popular communities. First of all this traumatizes anyone who gets to see it before the admins get to it, including the admins who have to review to take it down. Second of all, even if the content is a link to an external site, lemmy sill caches the thumbnail and stores it in the local pict-rs, causing headaches for the admins who have to somehow clear that out. Finally, both image posts and problematic thumbnails are federated to other lemmy instances, and then likewise stored in their pict-rs, causing such content to be stored in their image storage.
This has caused multiple instances to take radical measures, from defederating liberaly, to stopping image uploads to even shutting down.
Today I’m happy to announce that I’ve spend multiple days developing a tool you can plug into your instance to stop this at the source: pictrs-safety
Using a new feature from pictr-rs 0.4.3 we can now cause pictrs to call an arbitary endpoint to validate the content of an image before uploading it. pictrs-safety builds that endpoint which uses an asynchronous approach to validate such images.
I had already developed fedi-safety which could be used to regularly go through your image storage and delete all potential CSAM. I have now extended fedi-safety to plug into pict-rs safety and scan images sent by pict-rs.
The end effect is that any images uploaded or federated into your instance will be scanned in advance and if fedi-safety thinks they’re potential CSAM, they will not be uploaded to your image storage at all!
This covers three important vectors for abuse:
- Malicious users cannot upload CSAM to for trolling communities. Even novel GenerativeAI CSAM.
- Users cannot upload CSAM images and never submit a post or comment (making them invisible to admins). The images will be automatically rejected during upload
- Deferated images and thumbnails of CSAM will be rejected by your pict-rs.
Now, that said, this tool is AI-driven and thus, not perfect. There will be false positives, especially around lewd images and images which contain children or child-topics (even if not lewd). This is the bargain we have to take to prevent the bigger problem above.
By my napkin calculations, false positive rates are below 1%, but certainly someone’s innocent meme will eventually be affected. If this happen, I request to just move on as currently we don’t have a way to whitelist specific images. Don’t try to resize or modify the images to pass the filter. It won’t help you.
For lemmy admins:
- pictrs-safety contains a docker-compose sample you can add to your lemmy’s docker-compose. You will need to your put the .env in the same folder, or adjust the provided variables. (All kudos to @[email protected] for the docker support).
- You need to adjust your pict-rs ENVIRONMENT as well. Check the readme.
- fedi-safety must run on a system with GPU. The reason for this is that lemmy provides just a 10-seconds grace period for each upload before it times out the upload regardless of the results. A CPU scan will not be fast enough. However my architecture allows the fedi-safety to run on a different place than pictrs-safety. I am currently running it from my desktop. In fact, if you have a lot of images to scan, you can connect multiple scanning workers to pictrs-safety!
- For those who don’t have access to a GPU, I am working on a NSFW-scanner which will use the AI-Horde directly instead and won’t require using fedi-safety at all. Stay tuned.
For other fediverse software admins
fedi-safety can already be used to scan your image storage for CSAM, so you can also protect yourself and your users, even on mastodon or firefish or whatever.
I will try to provide real-time scanning in the future for each software as well and PRs are welcome.
Divisions by zero
This tool is already active now on divisions by zero. It’s usage should be transparent to you, but do let me know if you notice anything wrong.
Support
If you appreciate the priority work that I’ve put in this tool, please consider supporting this and future development work on liberapay:
All my work is and will always be FOSS and available for all who need it most.
Great work, this is the biggest issue that Lemmy has a the moment, I hope the admins will be able to set this up easily and start to take back all the preventative measures.
Sounds like progress, but please consider using a term other than “whitelist” when describing a list of allowed values. While the use of blacklist predates references to black as a race, allowlist is a reasonable alternative that doesn’t reinforce viewing black as less than or unwanted and white as allowed.
Sounds like progress, but please consider using a term other than “whitelist” when describing a list of allowed values. While the use of blacklist predates references to black as a race, allowlist is a reasonable alternative that doesn’t reinforce viewing black as less than or unwanted and white as allowed.
Holy fuck
Those are technical terms that have nothing to do with race or even humans.
Allowlist and Blocklist are also more intuitive to people who haven’t heard the terms before.
I’ve honestly always found “allowlist” and “blocklist” to feel like forced compound words, and I don’t see why “list” is necessary at all. For example, just saying “allowed” and “blocked” both implies it’s a list and is more intuitive than any of the *list terms.
Personally I have no stake in the battle, but I do wish people would use the most intuitive terms for the situation at least (whatever they are, for example “enabled”/“disabled” or “included”/“excluded”) instead of blanket ctrl+f on everything.
That’s a good point, and I hadn’t thought about that angle, that there just isn’t a reason for the terms to exist in the first place.
“In the red” and “in the black” is another pair that isn’t intuitive to me at all and I have to look up every time.
Oh the fuck with this nonsense!
Wahhhhhhh
Do you ever get tired of twisting yourself into a pretzel every time you want to be offended?
Making things that were never about race into things about race, just to have one more reason to be potentially offended by, is not productive and doesn’t help anyone.
By exercising enough mental gymnastics almost any term could be twisted into something supposedly offensive. The real solution to that problem: don’t do mental gymnastics.
Oh come on… The origin of blacklist was centuries before “black” became the term for a person of colour. And on a thread about CSAM…
Cool. After some testing Hexbear should run it. Not that the problem has ever been serious for Hex but still worthwhile and work that should absolutely be supported.
Unfortunately the GPU requirement is outside the scope of our current hosting and the nature of the material means it will not be hosted on a personal host
I think people are going to be much more concerned about the false negative rate than the false positive rate.
Sent you a little bit of money, @[email protected] - it sucks that this is necessary, but thanks for doing the good work <3
much appreciated
@[email protected]@[email protected]@[email protected] ok im gomna check this for wafrn because this was one of my initial fears when creating this
I’m curious how an AI like this is trained
it’s my understanding that the csam datasets (once already labeled by people) are hashed to the point of being unrecognizable before being passed around.
I don’t think training a model on hashes would be particularly useful - if the model were able to get any meaningful information out of it, that would mean the hash function itself is somehow leaking enough of the original contents to determine the image contents (which would essentially mean the hash function is broken beyond all repair)
deleted by creator
you know what that makes sense so I looked into it, one method of detecting csam is hashing the image and comparing it to a database of hashes of known csam images. so I guess that method might not work for “original” csam images.
apple csam detection [download]
interestingly, Google says they use AI also but didn’t really get into the details.
AFAIK it should work as long as the hashing function has some direct transformation property which can be extrapolated by the AI.
(Not an industry veteran, I dabble in AI mostly for hobby and sometimes work, but do have some accredited education on the subject. I may well be way off mark)
interesting read, thank you
given CLIP has a high zero-shot learning success rate, was it functional for this use case out of the box? or were further modifications required?
It requires specific usage of clip. Check the horde-safety repo if you’re interested
FWIW, the model they are using for this is not trained for CSAM detection at all. They are repurposing a Open AI tool called CLIP for a use case it was not made to support.
Edit: Not BLIP, CLIP
We’re not using blip
deleted by creator
Well done!
Great scott, you’ve done it!
Fantastic work
Fantastic work!
My first thought tho is of revenge. Is there any way to have it automatically report the incident to the FBI? Address of the uploader, etc
What if the troll is uploading from outside the USA ? I don’t think spaming authorities with reports will help…
You’re probably right
Also, if it reports a false positive, you’re sending an innocent person’s IP and other info to the authorities. I imagine the user won’t appreciate that.
Not unless lemmy and pict-rs devs provide a lot more functionality. However be aware that most hits will be false positives. You would be sending a lot of garbage over to the FBI unless you review every hit manually.
What about false positives?
Well someone would have to actually look at it if they want to prosecute
deleted by creator
False positive rate ~1 % False negative rate?
Invaluble work man, thanks!
holy hell this is massive
thank you for your work db0
