@wholesomedonut you're welcome. And thanks for the note; it helps to know that my reports are helping, because I wasn't sure if I should do that or just be silent and let it pass! Speaking of which, two new reports coming in 😛
@wholesomedonut@mike@fosstodon
Is it possible to collect their link/profile and share it? I'm thinking that it can be a nice project to use machine learning for spam detection. The only issue now is limited data I have.
We can build a tidy data out of these and then ask the community in a hackathon or fun competition fashion to come up with predictive models. This can be later published as a FLOSS tool for instance administration.
Would it be possible to discuss this matter? Are you folks interested at all?
This is a good opportunity to bring fediverse together and also get some media exposure behind it, unfortunately so far we have got ... well, see for yourself:
When deadline comes, we ask the teams to submit their model/method in form of XXX (either of docker, Nix shell, Guix shell, Apptainer, ... basically whatever the organizers of the event decides)
We run the submissions and test them on our internal test set (from step 3)
We announce the top 3 winners
Use one or combination of the models to assist instances in detecting spams (posts, accounts, ...) and flag them to admins.
This can be then become a FLOSS project, and with the help of fediverse platform developers be fitted as a plugin into them, or provide an API that these platforms can pass info to and get the "spam score".
Competition participants might even give SpamAssassin a try (although we are dealing with slightly different type of data than email).
@Mehrad Well, the data is pretty straight forward actually. All the accounts originated on instances with open registrations. The account names were randomized and exactly 10 characters each. There was no profile picture (/avatars/original/missing.png) and no header (/headers/original/missing.png). They followed nobody and had no followers (usually). They usually posted one of two images. Examples can be found with profile lm9ztxuz4q at instance don.neet.co.jp.
Yes, this is indeed one type of spam (actually a stupid one because a relatively simple query to database and some frequency analysis to confirm randomness could easily detect those.
But:
There are (and for sure will be) actual spams. Like accounts who spam hashtags or tag others and etc.
The media coverage portrait the recent issue as problematic and convoluted, and as illustration of Fediverse lack of spam detection
Tbh, at some point we will get more advanced spams, and if we don't have a oven-ready (as Boris called it 🤣) system to be deployed, what we get is bad publicity, frustrated admins and moderators, and annoyed users.
@thelinuxcast & @BrodieOnLinux were the first folks that I saw talking about spams as users, and @jerry and other frustrated admins and moderators.
But do you think there is no necessity for :air_quotes_left: Fedi Spam Assassin :air_quotes_right: ?
@Mehrad Absolutely there's a need of some way to recognize spam and deal with it. This spam wave was stupid. It wasn't subtle and how it was spreading was obvious, and we still barely had the tools required to keep it in check. It required hours of work from moderation teams. A ridiculous amount actually.
@mike
Many years back I submitted a research plan to Twitter to detect fake/spam/bot accounts based behavioral characteristics. They granted me developer access API for that and I pulled some data and did some work on it. It never materialized into a tangible result because I was also manually collecting data in form of lists to compare, and I never managed to have large enough cohort (twitter and people reporting was also shrinking my list).
But here we have a chance if we form teams I guess.
@wholesomedonut@mike@fosstodon
I've participated in multiple international ML competitions and I know a thing or two about the process and practices. Considering that Fediverse has high concentration of techy folks from different fields of expertise, I'd say it would be beneficial, fun and yet achievable to have a ML method for spam detection classification.
For the infrastructure, that would be easy to pull off as we want the ultimate model to be lightweight, fast, and deployable on VPS instead of beefy computer cluster. So we can allocate few normal computers (I have one spare Core i5) to run these models.
The models can come in form of docker, apptainer, nix, or guix containers (we should settle on one) and they can be run sequentially.
Add comment