Re: Reputation proxy

Simon Josefsson wrote:

> 
> Have you thought more about how submission into a reputation proxy would
> work?  That part is a bit sketchy to me, and the main problem seems to
> be how to avoid submissions from spammers.  Seed it with
> known-good-addresses, and then add new addresses based on who they
> communicate with?  There are also some serious privacy issues here.
> 
> I do like the general idea of a reputation proxy though.  It can be used
> for a lot of things beyond email, but perhaps the many small details
> makes it difficult to re-use one reputation proxy for another purpose.

sorry, I had a brain belch when I was writing.  I should have said proof of work 
as a proxy for the value of reputation.  In that a larger stamp conveys more 
reputation points on someone then a small or no stamp situation.  My reputation 
idea will probably work something like this.  Believe me, I am very open to 
hearing about flaws and ways of trying to fix it simply if possible.

in this context, reputation is a historical record of any given IP address.  It 
is aggregated into one of four or five bands.  The current thinking says five 
bands ranging from very bad to very good with neutral in the middle.  reputation 
is modeled on the human experience in that you earn a good reputation slowly but 
can lose it very quickly.

Reputation is driven by the messages received from an IP address.  Automatic 
processes such as filters on content give a small negative or positive boost to 
a reputation.  Proof of work puzzles at a medium positive boost to reputation. 
A human, sorting through the "unsure" collection, gives a strong positive or 
negative boost to reputation.

Reputation also modifies the filter configuration.  By themselves, a content 
filter or proof of work stamp is insufficient.  The filter is inadequate because 
of false positives and negatives, the proof of what works stamp is insufficient 
because a stamp large enough to defeat the super enormous bot net hypothesis 
consumes far more resources than any good mail site should spend.

I've tried a bunch of ways to describe the filter configurations so, please bear 
with me as I try one more.  Hopefully this will work.

As I said above, reputation chooses what filter is used to analyze the mail 
message.  There is no regular pattern.  Each configuration is admittedly 
subjective according to my biases but I believe each configuration is 
defensible.  This is not to say that other configurations wouldn't be just as 
valid and I would welcome a discussion of other configurations.

Configuration: very bad;  The filter configuration is a serial stamp, content 
filter.  That is to say, a message must pass the stamp challenge and the content 
filter in order to be delivered.messages from a very bad reputation source will 
never be delivered to an inbox.   Once the messages have passed the stamp/filter 
hurdle, messages that would have been delivered to the inbox or the spam trap 
are delivered to the spam trap.  Messages that score for delivery to the 
dumpster, are delivered the dumpster.  the presence of the stamp does not modify 
reputation in either direction.  Once a site is in the very bad state, the only 
reputation modifier is the spam trap processing i.e. messages trained as good 
improve reputation.

configuration: bad; the filter configuration is a mandatory content filter with 
optional stamp.  Like the very bad filter, messages are delivered only to the 
dumpster and spam trap destinations using the same rules.  If the stamp is 
present, the message will be delivered to the spam trap destination.

Configuration: neutral; mandatory content filter with optional stamp.
as messages come to the system, they're operated on by a variety of filters and 
user driven processes.  presence of stamp may or may not change message delivery 
from dumpster to spam trap.  Stamps at this reputation level do not enable 
direct to your inbox delivery.  They do however improve reputation at a rate 
faster than simple message exposure will.

Configuration: good; very good; delivery requires stamp or content filter. 
Stamped messages are delivered direct inbox, content filter messages are 
delivered to inbox or spam trap.  Dumpster scored messages are placed in a spam 
trap.  The only difference between good and very good reputation is the stamp size.

My rationale for these filters is admittedly subjective.  The very back 
configuration has two barriers to cross because the very bad category is 
effectively a blacklist.  I believe it is highly unethical to not have any way 
for a legitimate user to get through and scream for help or correction which is 
why the content filter plus stamp.  This also makes it damn near impossible for 
a spammer to get through as well.  A good user is more likely to have a simple 
message that will be recognized as good or as a spammer is more likely to have a 
message recognized as spam and therefore, fail of the tests

the bad category is my way of saying "I expect your message will be trash but, 
if you generate a stamp, it least it won't be thrown in the dumpster.  the 
filter only category is a concession to backwards compatibility but I've 
eliminated the inbox delivery range of messages because anything with a bad 
reputation doesn't deserved direct delivery.

The neutral category is simply filters as we know them today with ham, spam, and 
mystery meat categories (inbox, dumpster, spam trap).  The only thing a stamp 
gives you is faster movement from neutral to good reputation.

The good and very good categories are simply a safe place to enable direct 
delivery of messages in the presence of a stamp.  They also are safe place to 
eliminate the dumpster range for messages and either deliver to inbox or spam trap.


I'm afraid this is all I have time for today.  I look forward to your comments 
and questions.  I know I've left some important stuff out but I will deal with 
that in a later message.

---eric


-- 
Speech-recognition in use.  It makes mistakes, I correct some.