spam detection

[mayel]

I can think of three possible approaches to begin with:

• integration with spam detection API services #995
• document that akismet is an option and how to configure it
• add an optional ML model (running with NX) for detection -> needs testing and documentation
• Add per-user and per-instance keyword filtering #1001
• MRF for incoming federation Use MRF for spam detection #1049
• custom solution? (described below)

Custom solution could look like a simple scoring system, which adds/remove points based on things like:

• do I follow them?
• do they follow me, and if so how long
• do I follow someone who follows them, etc
• how many people are @ mentioned
• does it contain links
• have I mentioned them before
• is it a top-level post (vs a reply to a thread)
• etc

and if the score passes a configurable limit, do not actively notify the user (i.e. with push or email notification) and filter them out in separate tabs in notifications/messages views

could try making the algorithm using a configurable formula, so admins but even users can go in settings to improve the formula or simply tweak the sensitivity of different criteria in real time to better react to a spam influx, and share their algos or copy paste them from others... potential lib: https://hexdocs.pm/ex_pression/full_description.html

For inspiration see: pixelfed/pixelfed@dev/app/Util/Sentiment/Bouncer.php#L57-L100

[hellquist]

Have you considered knicking/replicating the MRF functionality of *oma services? It is extremely effective in stopping spam. There is a proposal to implement that for Mastodon, but I am thinking that the underlying thinking could be applied to any/all Fediverse services, and as I say, it has proven itself to really be working nicely (I run Akkoma and have had zero spam during this latest wave).

mastodon/mastodon#29252

[mayel]

Thanks for the suggestion! Luckily we do have MRF built-in: https://doc.bonfirenetworks.org/mrf.html (as our federation library started as fork of Pleroma). I will look into added the policies mentioned in that thread.