Hey folks!
I made a short post last night explaining why image uploads had been disabled. This was in the middle of the night for me, so I did not have time to go into a lot of detail, but I’m writing a more detailed post now to clear up where we are now and where we plan to go.
What’s the problem?
As shared by the lemmy.world team, over the past few days, some people have been spamming one of their communities with CSAM images. Lemmy has been attacked in various ways before, but this is clearly on a whole new level of depravity, as it’s first and foremost an attack on actual victims of child abuse, in addition to being an attack on the users and admins on Lemmy.
What’s the solution?
I am putting together a plan, both for the short term and for the longer term, to combat and prevent such content from ever reaching lemm.ee servers.
For the immediate future, I am taking the following steps:
1) Image uploads are completely disabled for all users
This is a drastic measure, and I am aware that it’s the opposite of what many of our users have been hoping, but at the moment, we simply don’t have the necessary tools to safely handle uploaded images.
2) All images which have federated in from other instances will be deleted from our servers, without any exception
At this point, we have millions of such images, and I am planning to just indiscriminately purge all of them. Posts from other instances will not be broken after the deletion, the deleted images will simply be loaded directly from other instances.
3) I will apply a small patch to the Lemmy backend running on lemm.ee to prevent images from other instances from being downloaded to our servers
Lemmy has always loaded some images directly from other servers, while saving other images locally to serve directly. I am eliminating the second option for the time being, forcing all images uploaded on external instances to always be loaded from those servers. This will somewhat increase the amount of servers which users will fetch images from when opening lemm.ee, which certainly has downsides, but I believe this is preferable to opening up our servers to potentially illegal content.
For the longer term, I have some further ideas:
4) Invite-based registrations
I believe that one of the best ways to effectively combat spam and malicious users is to implement an invite system on Lemmy. I have wanted to work on such a system ever since I first set up this instance, but real life and other things have been getting in the way, so I haven’t had a chance. However, with the current situation, I believe this feature is more important then ever, and I’m very hopeful I will be able to make time to work on it very soon.
My idea would be to grant our users a few invites, which would replenish every month if used. An invite will be required to sign up on lemm.ee after that point. The system will keep track of the invite hierarchy, and in extreme cases (such as spambot sign-ups), inviters may be held responsible for rule breaking users they have invited.
While this will certainly create a barrier of entry to signing up on lemm.ee, we are already one of the biggest instances, and I think at this point, such a barrier will do more good than harm.
5) Account requirements for specific activities
This is something that many admins and mods have been discussing for a while now, and I believe it would be an important feature for lemm.ee as well. Essentially, I would like to limit certain activities to users which meet specific requirements (maybe account age, amount of comments, etc). These activities might include things like image uploads, community creation, perhaps even private messages.
This could in theory limit creation of new accounts just to break rules (or laws).
6) Automated ML based NSFW scanning for all uploaded images
I think it makes sense to apply automatic scanning on all images before we save them on our servers, and if it’s flagged as NSFW, then we don’t accept the upload. While machine learning is not 100% accurate and will produce false positives, I believe this is a trade-off that we simply need to accept at this point. Not only will this help against any potential CSAM, it will also help us better enforce our “no pornography” rule.
This would potentially also allow us to resume caching images from other instances, which will improve both performance and privacy on lemm.ee.
With all of the above in place, I believe we will be able to re-enable image uploads with a much higher degree of safety. Of course, most of these ideas come with some significant downsides, but please keep in mind that users posting CSAM present an existential threat to Lemmy (in addition to just being absolutely morally disgusting and actively harmful to the victims of the abuse). If the choice is between having a Lemmy instance with some restrictions, or not having a Lemmy instance at all, then I think the restrictions are the better option.
I also would appreciate your patience in this matter, as all of the long term plans require additional development, and while this is currently a high priority issue for all Lemmy admins, we are all still volunteers and do not have the freedom to dedicate huge amounts of hours to working on new features.
As always, your feedback and thoughts are appreciated, so please feel free to leave a comment if you disagree with any of the plans or if you have any suggestions on how to improve them.
For step 6 - are you aware of the tooling the admin at dbzero has built to automate the scanning of images in Lemmy instances? It looks pretty promising.
Yep, I’ve already tested it and it’s one of the options I am considering implementing for lemm.ee as well.
It’s worth considering some commercially developed options as well: https://prostasia.org/blog/csam-filtering-options-compared/
The Cloudflare tool in particular is freely and widely available: https://blog.cloudflare.com/the-csam-scanning-tool/
I am no expert, but I’m quite skeptical of db0’s tool:
- It repurposes a library designed for preventing the creation of synthetic CSAM using stable diffusion. This library is typically used in conjunction with prompt scanning and other inputs into the generation process. When run outside it’s normal context on non-ai images, it will lack all this input context which I speculate reduces its effectiveness relative to the conditions under which it’s tested and developed.
- AI techniques live and die by the quality of the dataset used to train them. There is not and cannot be an open-source test dataset of CSAM upon which to train such a tool. One can attempt workarounds like extracting features classified and extracted separately like trying to detect coexisting features related to youth (trained from dataset A using non sexualized images including children) and sexuality (trained separately from dataset B using images containing only adult performers)… but the efficacy of open source solutions is going to be hamstrung by the inability to train, test, and assess effectiveness of the open tools. Developers of major commercial CSAM scanners are better able to partner with NCMEC and other groups fighting CSAM to assess the effectiveness of their tools.
I’m no expert, but my belief is that open tools are likely to be hamstrung permanently compared to the tools developed by big companies and the most effective solutions for Lemmy must integrate big company tools (or gov/nonprofit tools if they exist).
PS: Really impressed by your response plan. I hope the Lemmy world admins are watching this post, I know you all communicate and collaborate. Disabling image uploads is I think I very effective temporary response until detection and response tooling can be improved.
you make some good points. this gave rise to a thought: seems like law enforcement would have such a data set and seems they should of course allow tools to be trained on it. seems but who knows? might be worth finding out.)
It seems promising but also incomplete for US hosts, as our laws do not allow deletion of CSAM rather it must be saved and preserved and sent to a central authority and not deleted until they give the okay. Rofl.
I also wonder if this solution will use PHash or other hashing to filter out known and unaltered CSAM images (without actually comparing the images, rather their metadata).