[Rspamd-Users] Training rspamd

[Stefan Bauer]

Am Freitag, 17. April 2020 schrieb Kim Sindalsen via Users <
users at lists.rspamd.com>:

> Hi,
>
>
>
> Would it be possible to train the bayesian filter by setting up two emails
> -
> ie. ham at my.domain <mailto:ham at my.domain>  and spam at my.domain
> <mailto:spam at my.domain>  and then forward mails to those addresses?
>
>
>
> As in; make it train every message received at ham at my.domain as ham and
> vice
> versa?
>

We implemented something simililar. However keep the following in mind:

your users can not usually distinguish between spam and unwanted mails (
they might have subscribed for mailings previously and dont like it now)

you can pollute your filters if your users provide mails the wrong way

also make sure, you bypass all filters/milters for this kind of addresses

inform your users howto correctly forward mails in common mailclients, why
this is important and that there will be no confirmation to reports.

learning ham makes - for me - no sense. your users usually dont get any
false-positive spam mails to re-learn to there inboxes as you should
reject, what looks phishy

do not only learn 'spammy-mails' for bayes, also for fuzzy.

we do it the following way:

- maintain several spamtrap-addresses - mails to spamtraps are alway
learned with high spam score

have a 'spam-input' address that users can forward mails to an check this
mails by a human before learning it. do this asap as this prevents new
attacks to other users.

Stefan