Just for the record, here is my approach to my spam filter design just in case anyone is interested.
Design goals:
It should Just Work with an absolute minimum amount of user intervention and training required. That said, there are cases where user intervention will be necessary. In particular, emails that are sent in order to verify that the user controls their email address (password resets, 2FA auth codes, subscription confirmations) tend to be indistinguishable form phishing attacks and so some amount of user intervention will be needed.
It should be entirely server-side. This project was motivated in large part because my current spam filter is SpamSieve, which works great, but it runs on my laptop, so whenever my laptop is off-line so is my spam filter. So if I check my mail from my phone and my laptop is off-line I get a ton of spam.
I want to use as much off-the-shelf software as possible, but I also want to be able to have full control over the system. That means that whatever off-the-shelf software I use has to either Just Work, or be written in a language that I am proficient in. That rules out a lot of stuff because one of the languages I am *not* proficient in is Perl, and a lot of off-the-shelf spam filter stuff is written in Perl, and doesn’t do what I want out of the box.
Approach:
Seed the process with a source of reliable ham and reliable spam that does not require user labeling. The reliable ham is provided by keeping track of outgoing messages, which are presumed to be ham, and messages inbound to a honeypot address, which are presumed to be spam. The honeypot spam training corpus is shared among all users. The outgoing ham corpus is user-specific because who knows, someone may actually want information about Viagra.
The filter consists of two parts, a milter on the MTA side and a set of minimal sieve scripts on the LDA side that connect to a more or less traditional Bayesian filter. The milter tracks outgoing mail, and tags “easy” spam and ham on the incoming side. “Easy” ham consists of messages from senders to whom the user has previously sent messages, or with subjects that are “Re:” a subject about which the user has previously sent a message. Easy spam is things like Chinese text (at least that’s easy spam for me — that would obviously not work for a user in China, but this is mainly for my personal use) and certain super-spammy TLDs like .ru, .biz, etc. The milter also does greylisting. It’s written in Python using the pymilter library. The LDA-side Bayesian filter is written in Common Lisp. State is stored in a shared DB (currently SQLite3, but I’m probably going to switch to MariaDB because SQLite3 doesn’t play well with threads.)
What is left over after the milter is a set of messages from addresses with which the user has never corresponded. Those get put into an INCOMING folder and where they are processed by the Bayesian filter. The reason for putting the Bayesian filter on the LDA side is that this filter also applies the heuristic that if a message is received from two different unknown senders with the same subject within a relative short period of time (like 10-15 minutes) then both of those messages are almost certainly spam. Empirically, an MTA-side Bayesian filter misses a lot of easy spam because it cannot apply this heuristic. I mean yes, it’s possible, but it causes a lot of problems, not least of which is that timely things like password resets and 2FA auths get held up along with potential spam. (Greylisting has this problem too. I’m actually still trying to decide what to do about that.)
User training input is provided in the usual way, by having dovecot-sieve scripts that intercept messages being moved from INBOX to Junk and vice versa. (I have not yet decided what to do about messages that the user moves out of INCOMING.)
That’s where I’m at. Currently stuck on trying to figure out how to get the LDA-side filter to move messages. My baseline plan was to use external calls to doveadm, but it seems like there has to be a better way. Any and all advice and commentary much appreciated.
rg
On Jan 30, 2021, at 12:07 PM, Ron Garret ron@flownet.com wrote:
On Jan 30, 2021, at 11:54 AM, Tom Hendrikx tom@whyscream.net wrote:
IMHO you're still trying to re-invent the wheel :)
I don’t deny that. The goal of this project is as much (maybe more) to be a learning experience as it is to produce something useful.
FWIW, there are two reasons I don’t want to use a non-user-visible quarantine. First, there is always the possibility of a false positive, so all email must be made accessible to the user somehow. And second, there are occasions when you are expecting an email that looks spammy and you need to be able to get to it in a timely manner. The most common use case here is password reset links or 2FA authorization codes. It is not possible for a spam filter to distinguish a legitimate email of this type from a phishing attack. Only the user know if they recently requested a password reset. But *most* password reset emails are phishing attacks (at least most of the ones I get are) so I don’t want to see them by default.
rg