Re: [AMaViS-user] BAYES scoring is allowing more spam through?
On Sat, 07 Oct 2006, Jo Rhett wrote: 2. SA's self-registering stuff should be disabled - Bayes filters need to know what the human that is to receive the messages considers spam, not what a machine second-guesses. Thanks. I knew most of this, which is why I wasn't going to bother trying to train databases this time around. But what do you mean by self-registering ? That SpamAssassin enters messages with a certain low or high score into the Bayes filtering data bases automatically. ISTR SA calls this autolearning or something like that (bear with me, I don't use it). -- Matthias Andree - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Jo, Nope. I'm coming off of using CanIt, which does a damn good job. Since I can't afford the pro version for my personal machine, I was just checking out how good the open source interfaces have gotten. (besides spamassassin, which is obviously the core of CanIt's rules too) ... I do plan to look at SARE and sa-update and various other things, but right now I'm working deliberately with as-bone-stock-as-possible just to see how well it works out of the box. As an experiment it is valid to try bare-bones SA, although for production use most sites use some subset of SARE rules, and with more recent versions of SA the use of sa-update is very much recommended, as it adds a couple of very useful last-minute additions or fixes to base rules. It would not be fair to judge SA without SARE, sa-update (and network test and bayes). I'm fairly pleased with amavis, but struggling with the lack of documentation. Something that I plan to spend a lot of November improving, once I've gotten my head around all of it. Patrick Ben Koetter p at state-of-mind.de is investing his time in documenting amavisd-new, the project is progressing steadily. If you have serious intentions in helping with documentation, please contact him to avoid duplicating efforts. Mark - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Mark Martinec wrote: As an experiment it is valid to try bare-bones SA, although for production use most sites use some subset of SARE rules, and with more recent versions of SA the use of sa-update is very much recommended, as it adds a couple of very useful last-minute additions or fixes to base rules. It would not be fair to judge SA without SARE, sa-update (and network test and bayes). I had come to that conclusion myself, but only by research. This should be documented better (which leads to the next point...) Patrick Ben Koetter p at state-of-mind.de is investing his time in documenting amavisd-new, the project is progressing steadily. If you have serious intentions in helping with documentation, please contact him to avoid duplicating efforts. Will do. Thanks. -- Jo Rhett Network/Software Engineer Net Consonance - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Jo Rhett [EMAIL PROTECTED] writes: I keep seeing spam messages with negative scores, like this: No, score=-0.317 tagged_above=-1.99 required=4.01 tests=[BAYES_00=-2.599, DNS_FROM_RFC_ABUSE=0.2, DNS_FROM_RFC_POST=1.708, HTML_30_40=0.374, HTML_MESSAGE=0.001, MIME_HTML_ONLY=0.001, NO_RECEIVED=-0.001, NO_RELAYS=-0.001] So I'm trying to figure out what amavisd is doing with the sa bayes stuff, but all I can find is a few notes that amavisd-new uses a single bayes database for all users. Nothing about how to tweak it ... Take this with a grain of salt, I'm biased (being the bogofilter co-maintainer): IMHO the SpamAssassin Bayes stuff should be disabled -- for various reasons: 1. is that it causes a major performance issue, even on lightly loaded Pentium D and Xeon 2.8 equipped servers, which causes filtering to be aborted through amavis's timeouts (at least for my older amavisd-new version). 2. SA's self-registering stuff should be disabled - Bayes filters need to know what the human that is to receive the messages considers spam, not what a machine second-guesses. 3. centralistic Bayes filtering doesn't work well, as you've just seen in one of the failure modes that are sort of expected for this class of filters 4. per-user databases don't scale too well 5. Bayes requires users to be permanently alert and correct mistakes. I'm not saying that SpamAssassin alone or Bayes are bad, but SpamAssassin's Bayes implementation is, and integrating it with amavisd-new doesn't ameliorate the pain. 3 to 5 are generic issues that affect all trainable/Bayesian filters. -- Matthias Andree - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Gary V wrote: It's sad, but we are at war with the spammers and the virus writers and the adware companies and the hackers and the spyware writers. Their tactics change on a daily basis, and so must ours. War often requires effort, sorry. I know that. Which is why I keep pointing out that spammers have adopted tactics to deliberately poison bayes caches. But everyone ignores that. Unless the bayes mechanisms are updated to be smarter, the current crop of bayes-polluting auto-bot spam will invalidate everyone's bayes. What version of SpamAssassin are you running? Are you using any SARE rules? Nope. I'm coming off of using CanIt, which does a damn good job. Since I can't afford the pro version for my personal machine, I was just checking out how good the open source interfaces have gotten. (besides spamassassin, which is obviously the core of CanIt's rules too) I'm fairly pleased with amavis, but struggling with the lack of documentation. Something that I plan to spend a lot of November improving, once I've gotten my head around all of it. I do plan to look at SARE and sa-update and various other things, but right now I'm working deliberately with as-bone-stock-as-possible just to see how well it works out of the box. -- Jo Rhett Network/Software Engineer Net Consonance - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Matthias Andree wrote: Take this with a grain of salt, I'm biased (being the bogofilter co-maintainer): IMHO the SpamAssassin Bayes stuff should be disabled -- for various reasons: 1. is that it causes a major performance issue, even on lightly loaded Pentium D and Xeon 2.8 equipped servers, which causes filtering to be aborted through amavis's timeouts (at least for my older amavisd-new version). 2. SA's self-registering stuff should be disabled - Bayes filters need to know what the human that is to receive the messages considers spam, not what a machine second-guesses. 3. centralistic Bayes filtering doesn't work well, as you've just seen in one of the failure modes that are sort of expected for this class of filters 4. per-user databases don't scale too well 5. Bayes requires users to be permanently alert and correct mistakes. I'm not saying that SpamAssassin alone or Bayes are bad, but SpamAssassin's Bayes implementation is, and integrating it with amavisd-new doesn't ameliorate the pain. 3 to 5 are generic issues that affect all trainable/Bayesian filters. Thanks. I knew most of this, which is why I wasn't going to bother trying to train databases this time around. But what do you mean by self-registering ? -- Jo Rhett Network/Software Engineer Net Consonance - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
On Sat, Oct 07, 2006 at 09:58:49AM -0700, Jo Rhett wrote: Matthias Andree wrote: Take this with a grain of salt, I'm biased (being the bogofilter co-maintainer): IMHO the SpamAssassin Bayes stuff should be disabled -- for various reasons: 1. is that it causes a major performance issue, even on lightly loaded Pentium D and Xeon 2.8 equipped servers, which causes filtering to be aborted through amavis's timeouts (at least for my older amavisd-new version). 2. SA's self-registering stuff should be disabled - Bayes filters need to know what the human that is to receive the messages considers spam, not what a machine second-guesses. ... Thanks. I knew most of this, which is why I wasn't going to bother trying to train databases this time around. But what do you mean by self-registering ? autolearn - The default is for SA to feed into Bayes anything that scores high enough (as spam) or low enough (as ham). I agree that this is a fundamentally and philosophically flawed idea. At the least it should not be a default and should use much further thresholds than it did when I last checked. If you weren't aware of this, this may be how your Bayes DBs ended up poisoned. With autolearn it can happen anytime a new style of spam comes through which is not caught by SA's existing rules. -- Clifton -- Clifton Royston -- [EMAIL PROTECTED] / [EMAIL PROTECTED] President - I and I Computing * http://www.iandicomputing.com/ Custom programming, network design, systems and network consulting services - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Jo wrote: Gary V wrote: It's sad, but we are at war with the spammers and the virus writers and the adware companies and the hackers and the spyware writers. Their tactics change on a daily basis, and so must ours. War often requires effort, sorry. I know that. Which is why I keep pointing out that spammers have adopted tactics to deliberately poison bayes caches. But everyone ignores that. I would say that because of the way Bayes works, this does happen, but also because of the way Bayes works, it can be prevented. I appreciate the fact you have illustrated that if one chooses not to maintain their database, this may result. Unless the bayes mechanisms are updated to be smarter, the current crop of bayes-polluting auto-bot spam will invalidate everyone's bayes. This could be rephrased: Unless the anti-spam mechanisms are updated to be smarter, the current crop of spam will invalidate everyone's anti-spam mechanisms. Not surprising then that a new version of SpamAssassin comes out so often, and rules (SA or 3rd party) change on a nearly daily basis. Personally, my only goal is to attempt to solve problems with what is currently available. For your system, if your Bayes is corrupt, then work toward setting it right or get rid of it. SpamAssassin will work without it, and is designed to compensate for the loss. Gary V - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Jo, Now, the spammers are putting lots of junk text in their spam and polluting the databases to such an extend that Bayes is much less useful. So I guess I'm saying that I have very little interest in spending the effort to retrain a new Bayes database, and none of my other users are capable or clueful enough to do so. I hardly ever need to train bayes (1000 users, an organization, not an ISP), I just feed it half a dozen spam messages per week that got through. It is essential that your other rules are good, including dcc, razor, uribls, sa-update rules, SARE rules, FuzzyOCR and possibly a handful of custom rules. Also p0f rules help. So other rules apply their collective knowledge to auto-train bayes, and bayes pays back with its digested knoweledge, like a large flywheel. Mark - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Okay, so I used to deal with Bayes quite a bit. I spent a very long time specially training my Bayes database, and it seemed to work. Now, the spammers are putting lots of junk text in their spam and polluting the databases to such an extend that Bayes is much less useful. So I guess I'm saying that I have very little interest in spending the effort to retrain a new Bayes database, and none of my other users are capable or clueful enough to do so. Short of disabling Bayes entirely, is there anything I can do with minimal effort that yields a minimal return? I'm just not convinced that Bayes databases are going to stack up well against the modern bayes-scatter spam. On Oct 5, 2006, at 10:32 PM, Gary V wrote: If your effort was some time ago, those tokens are long gone by now. The rate at which new tokens are added is somewhat the same as the ... If you quarantine, consider feeding all the spam in your quarantine to Bayes (after you remove any false positives). Sometimes spam like the .. deletes everything once it is learned. I admit I don't do it every day, but at least two or three times a week. I also have a script that deletes anything older than 30 days from both accounts - simply so the server won't croak if I should happen to. :) I really, really don't want to be rude but who are you replying to? You apparently didn't read a single word of what I wrote above. Really, not trying to be rude -- just can't follow this thread. NO. I won't spend another minute of my day training Bayes, and nobody else on this server is clueful enough. Period, end of subject. Done. Now, how can I prevent Bayes from SUBTRACTING 2.6 from every message short of completely disabling it? -- Jo Rhett Senior Network Engineer Network Consonance - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
On Oct 6, 2006, at 1:26 AM, Mark Martinec wrote: I hardly ever need to train bayes (1000 users, an organization, not an ISP), I just feed it half a dozen spam messages per week that got through. It is essential that your other rules are good, including dcc, razor, uribls, sa-update rules, SARE rules, FuzzyOCR and possibly a handful of custom rules. Also p0f rules help. So other rules apply their collective knowledge to auto-train bayes, and bayes pays back with its digested knoweledge, like a large flywheel. Okay, that's good. Now, how can I avoid having Bayes subtract 2.6 from every message until it gets a clue? Also, are there any commands to see what bayes knows about, thinks about, etc? -- Jo Rhett Senior Network Engineer Network Consonance - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
On Fri, 6 Oct 2006 10:44 -0700, Jo Rhett wrote: Now, how can I prevent Bayes from SUBTRACTING 2.6 from every message short of completely disabling it? These lines in /usr/local/etc/mail/spamassassin/local.cf (or whatever path your local.cf is in) and then amavisd reload should do it I think: score BAYES_00 0 0 0 0 score BAYES_05 0 0 0 0 score BAYES_20 0 0 0 0 score BAYES_40 0 0 0 0 Those four categories are the ones with negative values as default. See grep BAYES /usr/local/share/spamassassin/50_scores.cf. Peter Olsson[EMAIL PROTECTED] - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Jo wrote: Also, are there any commands to see what bayes knows about, thinks about, etc? This one can (at least) show number of learned spam and ham: su vscan -c 'sa-learn --dump magic' 0.000 0 158089 0 non-token data: nspam 0.000 0 19527 0 non-token data: nham Gary V - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Jo wrote: I really, really don't want to be rude but who are you replying to? You apparently didn't read a single word of what I wrote above. Really, not trying to be rude -- just can't follow this thread. On Fri, October 6, 2006 11:01 am, Gary V wrote: It is my opinion that only people who feel they are need to say stuff like that. Yes exactly so. If I had known you better, I would have been rude :-) But without context and being new on the list, you wouldn't know me well enough to know I'm joking. People can take things far too seriously ;-) Now, how can I prevent Bayes from SUBTRACTING 2.6 from every message short of completely disabling it? adjust scores, here are likely current settings: This I knew already. I was questioning if doing so would make bayes invalid enough that I should simply disable Bayes entirely? There might be more and better logic for this... - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Jo wrote: I really, really don't want to be rude but who are you replying to? You apparently didn't read a single word of what I wrote above. Really, not trying to be rude -- just can't follow this thread. On Fri, October 6, 2006 11:01 am, Gary V wrote: It is my opinion that only people who feel they are need to say stuff like that. Yes exactly so. If I had known you better, I would have been rude :-) But without context and being new on the list, you wouldn't know me well enough to know I'm joking. People can take things far too seriously ;-) Now, how can I prevent Bayes from SUBTRACTING 2.6 from every message short of completely disabling it? adjust scores, here are likely current settings: This I knew already. I was questioning if doing so would make bayes invalid enough that I should simply disable Bayes entirely? There might be more and better logic for this... - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
jrhett wrote: Now, how can I prevent Bayes from SUBTRACTING 2.6 from every message short of completely disabling it? adjust scores, here are likely current settings: This I knew already. I was questioning if doing so would make bayes invalid enough that I should simply disable Bayes entirely? There might be more and better logic for this... Adjusting Bayes scores does not affect Bayes. Bayes is not recursive. If you know all the stuff that has been discussed, then you know enough about Bayes to realize that it needs good data to make decisions. A lot of stuff can be auto-learned, but when you find spam like the one that scored BAYES_00, then you need to manually feed it to Bayes. You desire to not spend any time doing that (none of us do), but I don't see how you can correct your out-of-whack (possibly poisoned) database without putting *some* sort of effort into retraining it (even if that means dumping it and starting over - and letting it simply autolearn - which is a valid option for a poisoned database). It's sad, but we are at war with the spammers and the virus writers and the adware companies and the hackers and the spyware writers. Their tactics change on a daily basis, and so must ours. War often requires effort, sorry. Mark's suggestions are good ones: make sure everything else is working as good as it can so it is more likely spam will score over your autolearn threshold. I also suggested lowering your thresholds. The only harm that zeroing out your lower scoring Bayes rules will cause is your legitimate mail is likely to score higher. Whether it makes a difference or not depends on your tag2_level, kill_level and in general, what typical legitimate mail is currently scoring at. You should probably keep an eye on it if you zero them out. What version of SpamAssassin are you running? Are you using any SARE rules? Gary V - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Jo wrote: I keep seeing spam messages with negative scores, like this: No, score=-0.317 tagged_above=-1.99 required=4.01 tests=[BAYES_00=-2.599, DNS_FROM_RFC_ABUSE=0.2, DNS_FROM_RFC_POST=1.708, HTML_30_40=0.374, HTML_MESSAGE=0.001, MIME_HTML_ONLY=0.001, NO_RECEIVED=-0.001, NO_RELAYS=-0.001] So I'm trying to figure out what amavisd is doing with the sa bayes stuff, but all I can find is a few notes that amavisd-new uses a single bayes database for all users. Nothing about how to tweak it ... This is a SpamAssassin issue, but generally it appears your Bayes could use some manual training: http://spamassassin.apache.org/full/3.1.x/dist/doc/sa-learn.html Run sa-learn as the amavisd-new user, e.g. : su vscan -c 'sa-learn --spam /path/to/spam' If the spam is not on your amavisd-new system, see if you can save the spam as plain text files (that are readable as such) and transfer them to your amavisd-new system using something like WinSCP. Having the mail in a format usable by sa-learn is a frequent topic on the SA list. It's not an issue when the mail store is on the same server as SpamAssassin, but often difficult when it is not. For example: http://wiki.apache.org/spamassassin/SiteWideBayesFeedback Gary V - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Gary V wrote: This is a SpamAssassin issue, but generally it appears your Bayes could use some manual training: http://spamassassin.apache.org/full/3.1.x/dist/doc/sa-learn.html Okay, so I used to deal with Bayes quite a bit. I spent a very long time specially training my Bayes database, and it seemed to work. Now, the spammers are putting lots of junk text in their spam and polluting the databases to such an extend that Bayes is much less useful. So I guess I'm saying that I have very little interest in spending the effort to retrain a new Bayes database, and none of my other users are capable or clueful enough to do so. Short of disabling Bayes entirely, is there anything I can do with minimal effort that yields a minimal return? I'm just not convinced that Bayes databases are going to stack up well against the modern bayes-scatter spam. -- Jo Rhett Network/Software Engineer Net Consonance - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/
Re: [AMaViS-user] BAYES scoring is allowing more spam through?
Jo wrote: Gary V wrote: This is a SpamAssassin issue, but generally it appears your Bayes could use some manual training: http://spamassassin.apache.org/full/3.1.x/dist/doc/sa-learn.html Okay, so I used to deal with Bayes quite a bit. I spent a very long time specially training my Bayes database, and it seemed to work. Now, the spammers are putting lots of junk text in their spam and polluting the databases to such an extend that Bayes is much less useful. So I guess I'm saying that I have very little interest in spending the effort to retrain a new Bayes database, and none of my other users are capable or clueful enough to do so. Short of disabling Bayes entirely, is there anything I can do with minimal effort that yields a minimal return? I'm just not convinced that Bayes databases are going to stack up well against the modern bayes-scatter spam. If your effort was some time ago, those tokens are long gone by now. The rate at which new tokens are added is somewhat the same as the rate that old tokens are discarded. I assume you are auto-learning. It looks like if nothing else, you might have to feed stuff like the spam in question to Bayes. You can adjust the auto-learn thresholds down, this might help a little, possibly something like: bayes_auto_learn_threshold_nonspam -0.5 bayes_auto_learn_threshold_spam 9.0 If you quarantine, consider feeding all the spam in your quarantine to Bayes (after you remove any false positives). Sometimes spam like the one in question will score low on Bayes, but high enough on other stuff to get quarantined. I have a relay server but I use the virtual map to fork a copy of my personal mail to local Maildir style folders. I also quarantine locally to a mail box and use Courier IMAP to access these two accounts. I move spam in my account to a spam folder and check the quarantine for false positives (which almost never happens). Once this is handled, I run a script to learn everything - ham and spam. Everything in both accounts is disposable at this point, so my script also deletes everything once it is learned. I admit I don't do it every day, but at least two or three times a week. I also have a script that deletes anything older than 30 days from both accounts - simply so the server won't croak if I should happen to. :) Gary V - Take Surveys. Earn Cash. Influence the Future of IT Join SourceForge.net's Techsay panel and you'll get the chance to share your opinions on IT business topics through brief surveys -- and earn cash http://www.techsay.com/default.php?page=join.phpp=sourceforgeCID=DEVDEV ___ AMaViS-user mailing list AMaViS-user@lists.sourceforge.net https://lists.sourceforge.net/lists/listinfo/amavis-user AMaViS-FAQ:http://www.amavis.org/amavis-faq.php3 AMaViS-HowTos:http://www.amavis.org/howto/