On Mon, Sep 19, 2022 at 3:16 PM Slavko via mailop <[email protected]> wrote:
> Dňa 19. septembra 2022 21:44:08 UTC používateľ Brandon Long via mailop < > [email protected]> napísal: > > >Using machine learning for personal mail or very few users, not sure if > >that's likely to > >be worth the investment unless you want to learn a lot about it. You are > >likely to have > >a good history of "good" mail in an archive, at least, which means a model > >to learn what > >to whitelist may work well. > > While i build my own archive of SPAMs, it is still small amount of messages > in numbers. I do not access delivered messages without exact user's > acknowledge (despite that i can), but that are small numbers too. Thus > after initial experiments, i abandon the mentioned neural (and fuzzy) > rspamd's modules as wasting of resources (at least in my case). > > Would be great, if some bigger providers can share their models after > training, but that again comes with problems, first is trust, then updates, > and (of course) forrmat, thus IMO that cannot work, at least not yet. But > perhaps it will evolve to something widely usable (as RBLDNS are) in > future. > > Models are highly dependent on the features (signals) they use, and computing those features would also need to be opened with the model... and it would make it easier for spammers to design email to avoid the model, somewhat. In our case, the model also highly relies on a bunch of other systems such as the reputation for particular email features such as IPs and domains. There is also the general question of whether a given model contains "PII" because it was trained on user PII. Publishing models is complicated. Brandon
_______________________________________________ mailop mailing list [email protected] https://list.mailop.org/listinfo/mailop
