Phill Macey a écrit : > 2009/11/7 mouss <mo...@ml.netoyen.net>: >> >> Most statistical anti-spam filters assume an inbound model. you can use >> a "global" bayes setup, but then I don't think you'll benefit from >> dspam/bogo/... >> > > Could you turn the outgoing mail around and make it inbound mail as > well? [snip]
no, the problem is related to training. in the case of inbound mail, statistical filters use the fact that a given user (or a given set of users) receive mail which characteristics can be learned if you have a sufficient corpus (of ham and spam). you can still use this for outbound mail, with a global "dictionary" (site wide setup). but - nobody is going to feed back "false negatives" (missed spam) - who is going to feed back "false positives"? how? while feasible, this is not a simple problem. that said, you can still run spamassin and have a log parser to detect problems: some user suddenly sends a lot of mail that gets tagged as spam... etc. definitely not a simple problem...
