Future Project Plans and Ideas
Last Updated: Oct 15, 2004

1. Neural networking should probably be rewritten; the experimental version
   in DSPAM works, but is not truly "neural" in that each node doesn't
   recursively process its own interconnections, but all interconnections are
   processed in parallel (making it very unusable for very large groups). I'm 
   not entirely sure if neural networking is the right term either, as we're 
   not really "training" the network per se; we are calculating an answer to 
   which we do not know the answer to (the N most reliable nodes to a user).

   The ideal redesign for this would include a set of interconnections to the
   input node which are processed first, and then depending on weighting the
   interconnections of the child nodes should be processed, and so on. The
   output would be the N most reliable nodes for a particular user as
   determined based on the reliability of the child and sub-child nodes.

   A quick explanation of the goal here: to find other users in the system who
   can provide the most reliable council when the filter is uncertain about
   a message.

2. Another type of neural networking I am considering could be called something
   like "associative neural inference". The idea behind this is to track
   interconnections between adjacent words, like chained tokens, only use
   these interconnections to instantiate tokens which aren't present in a
   message. For example if "free", "special", and "great" are present in an
   email and all have a very strong interconnection with the word "offer"
   (which isn't present), the word offer would be instantiated as a token as
   if it were in the email. This will give DSPAM another layer of AI in that
   it will begin to "guess" what the context is about.

3. Internationalization - Still haven't worked with this before. Need to add
   support for wide character emails. User should have the ability to turn
   this on/off, as having wide chars Z'd out makes foreign spams easier to
   identify.

4. Language packs. The DSPAM agent is pretty easy to configure for different
   languages, as all the basic dialog is in language.h. The CGI should have
   support for language packs, a single file containing a key to different
   text, and then code to interpolate those texts in the templates. 

5. dspamc LMTP authentication. When using dspamc, I don't like authenticating 
   using MAIL FROM. With LMTP, we should be able to implement SMTP AUTH.

What it looks like now:

SERVER> 220 DSPAM DLMTP 3.4.0 Authentication Required
CLIENT> LHLO Relay1
SERVER> 250-PIPELINING
SERVER> 250-ENHANCEDSTATUSCODES
SERVER> 250-DSPAMPROCESSMODE
SERVER> 250 SIZE
CLIENT> MAIL FROM: <secret@Relay1> DSPAMPROCESSMODE="--deliver=innocent -d %u"
SERVER> 250 2.1.0 OK
CLIENT> RCPT TO: jonz
SERVER> 250 2.1.5 OK
CLIENT> DATA
SERVER> 354 Enter mail, end with "." on a line by itself
CLIENT> Subject: Cheap Viagra!
CLIENT>
CLIENT> Click Here: http://www.cheapviagra.com
CLIENT> .
SERVER> 250 2.0.0 <jonz> Message accepted for delivery: INNOCENT

What it SHOULD look like:

SERVER> 220 DSPAM LMTP 3.4.0 Ready
CLIENT> LHLO Relay1
SERVER> 250-PIPELINING
SERVER> 250-ENHANCEDSTATUSCODES
SERVER> 250-DSPAMPROCESSMODE
SERVER> 250-SIZE
SERVER> 250 AUTH DIGEST-MD5
CLIENT> AUTH DIGEST-MD5
SERVER> 334
SERVER> PENCeUxFREJoU0NnbmhNWitOMjNGNndAZWx3b29kLmlubm9zb2Z0LmNvbT4
CLIENT> ZnJlZCA5ZTk1YWVlMDljNDBhZjJiODRhMGMyYjNiYmFlNzg2ZQ==
SERVER> 235 Authentication Successful
CLIENT> MAIL FROM: <> DSPAMPROCESSMODE="--deliver=innocent -d %u"
SERVER> 250 2.1.0 OK
CLIENT> RCPT TO: <jonz>
SERVER> 250 2.1.5 OK
CLIENT> DATA
SERVER> 354 Enter mail, end with "." on a line by itself
CLIENT> Subject: Cheap Viagra!
CLIENT>
CLIENT> Click Here: http://www.cheapviagra.com
CLIENT> .
SERVER> 250 2.0.0 <jonz> Message accepted for delivery: INNOCENT