Re: PR Problem?

Mario 'BitKoenig' Holbe wrote:
> Eric S. Johansson <[email protected]> wrote:
>> DeLesley SpamBox wrote:
>>> I'm not convinced that even a naive sender pays wouldn't be helpful.
>> make assumptions about the number of zombies, how much leakage you will 
>> permit and you can get the stamp size.  It's quite entertaining.
> 
> I don't think this is a good argument. Even now/today all these zombies
> could be used to generate spam mails directly. So the question should
> rather be: how far would the amout of spam messages decrease when all
> these zombies would additionally need to pay CPU for sender stamps.

stamps work in three ways in this context.  As the load on the spammer, 
as an indicator of the quality of the e-mail source, and as an introducer.

The load on the spammer is self-evident.  The more work they have to do 
to generate each piece of spam, the lower the profit margin by virtue of 
the lowered visibility of their traffic.  The quality of the e-mail 
source is only apparent if you have a feedback mechanism between sender 
and receiver of e-mail.  If I get spam from a variety of addresses, the 
cost of getting e-mail from that address goes up.  I can also increase 
costs by specifying always stamp versus first time stamp.  But like I 
said, this only works if the stamper queries the recipient 
domains/domains  to find out how much postage should be.  See previous 
conversation about DNS/HTTP-based mechanisms for communicating postage 
costs.  The introducer model says that a stamp introduces you to 
somebody else as a way of bypassing filters.  This allows you to bias 
your spam filter scores so that those with stamps get through in 
preference to those without.  This may or may not be a good thing but I 
have been known to be a little biased at times.

>> you are falling into the classic trap of assuming that the cost of 
>> hardware mean something.  This is the fallacy behind the Ben Laurie 
>> paper.  It's important to remember that the cost per stamp drops with 
>> every stamp generated with a given piece of hardware.  The first step is 
> 
> I didn't read the Ben Laurie paper, if I should do, because it proves my
> below aguments wrong, please tell me :)
> Of course, sender stamps can only reduce the total amount of spam by a
> linear factor. A big linear factor probably, but linear.
> However, the nice thing about sender stamps is that this linear factor
> is very easy adjustible to the average current hardware out there. And
> this is why hardware costs begin to mean something.

stamp "size" shifts with time as hardware performance improves.  But if 
you lay out $1000 for a fast machine, every stamp you calculate drops 
the cost per stamp.  Ranging from a very expensive first stamp to the 
number stamps per day * 365 Which means stamps are frighteningly cheap 
if you use that metric.  Even with power and cooling averaged in, it's 
still amazingly cheap.  As a friend said, how expensive can it be to run 
a dozen PCs in a back room in Jakarta with the fan blowing in "cooler" air.

> By just calculating the "average price" (quantiled average over the size
> of stamps - quantiled to prevent DoS) of all emails you get, your MUA
> can easily find out how much it *needs* to pay for the stamp to get a
> good probability for the delivery of your mails. Of course, it can
> always calculate bigger stamps, if it or his user likes. By using a
> min() function over the above average and what the MUA is able to
> calculate within a user-defined time, the above average slides over
> time and thus adjusts to the average hardware out there and to what
> users are willing to pay.

> On the other side, MTAs, i.e. spam filters, can adjust their price-
> acceptance function for sender stamps that simple as well.

this is why the feedback mechanism proposed is a good idea.  You don't 
need to guess the cost of a stamp you just do it.  In fact, the 
technique described as the vantage that it gradually renders all of the 
spam generating zombie addresses effectively useless.  This is not to 
say they can't deliver spam, it's just that it costs a whole lot to do so.


> Of course the average price is hard for PDAs, slow machines etc.
> However, at any time MTAs can calculate sender stamps theirselves on
> behalf of the sender (as sendmail-hashcash shows). So MTAs could
> easily generate sender stamps for authenticated well-known clients.
> Of course, the best solution would be some incremental algorithm, where
> you can subsequently increase the stamp size just by investing a bit
> more CPU time.

if you use the appropriate variable postage mechanism, once you 
establish communications with someone, postage ceases to be an issue. 
This becomes a freebie given to you by the service provider because 
you're paying them $30, $40, $75, or more a month.  they can afford to 
spend some of the money you leave on the table for the five or 10 stamps 
a month you'll need.  get you also have the option of saying "I'll take 
my chances" and not send a stamp.

> 
>> stay in business.  The number of zombies will decrease and be more 
>> easily targeted.
> 
> Well, then users need to be willing to pay more for their own stamps.

if you pay attention to basic human factors, sure, they will.  Generate 
and background, the user doesn't see anything and if you use variable 
postage mechanisms, the number stamps per user per day will be trivially 
small.


>> this is probably a philosophical disagreement.  I absolutely of abhor 
>> false positives.  I look in the dumpster maybe once every couple of 
>> months if somebody tells me something was lost.  I look in my spam trap 
>> about once a week.  If somebody is going to send me a message with a 
>> stamp, I have no problem with it coming through directly.  If it's a 
>> spammer, I want to be able to mark it as spam and then permanently 
>> blacklisted IP address and tell all of my friends about it automatically.
> 
> Well, I personally think this is a bit a blue-eyed point-of-view. If you
> think this is really feasible, just think about why you don't just do
> the same today without stamps.
> The more stamps become widely accepted, the more spammers will use them
> as well. And... wasn't this the idea anyways? Spammers should be forced
> to use them to increase the cost for spam :)

well, it's how I live and without stamps.  It's been rather successful I 
might add.  Anyone who uses twopenny blue also basically lives the same 
way.  if I didn't have this ability to not look at my spam trap, I 
probably would have ditched e-mail long time ago for something more 
useful like the telephone.

but the addition of stamps as filter bypass improves the quality of 
system behavior because it now becomes predictable.  It used to be the 
e-mail was predictable, for the most part, in that it either was 
delivered or it wasn't.  It was so reliable that people didn't care 
about the unreliability warnings, it just worked.  But now with spam and 
the probabilistic content filters, e-mail has become unreliable because 
it's unpredictable.  You have no idea if a customer's spreadsheet with 
HTML framework talking about the shipment of strawberries out of 
California is going to pass your content filter or not.  This is 
unacceptable for businesses.

Lest you think I'm making this scenario up, this is one of many I've 
lived with one of my customers, a fruit and vegetable wholesaler.  Their 
salespeople send invoices, quotes etc. and receive the same by e-mail. 
If e-mail is down, you can hear the thousand dollar counter clicking 
rather rapidly in the background.  An e-mail lost in a spam trap for a 
day can literally cost them tens of thousands of dollars.  And this is a 
small operation.  If you want e-mail to become reliable again, you need 
a predictable and determine a deterministic event which says "this stamp 
will get through".  If there is some way to combine the two models, I'm 
open.  Let's see what can fly for those who don't really care if a 
message arrives versus those that do.

>> to use a stamp or even a stamp size as a scoring factor actually works 
>> in the spammers favor.  By crafting a message the right way and just 
>> putting a little stamp, maybe 10 seconds worth, they would be able to 
>> almost guaranteed delivery.  While at the same time, you would still end 
> 
> Hehe, so there are methods out there to reduce the amount of work that
> is needed to calculate a stamp? :)
> If not: the automatic adaption of MUAs and MTAs to the stamp size works
> against spammers using too small stamps.

I was concerned that stamps as a modifier for content filter scores 
could give spammers a leg up at making their messages more visible for 
very little work.  It's very simple to analyze the reduction of spam on 
the net using a stamp as filter bypass.  You can only know what effect a 
small stamp would have in conjunction with a filter if you looked at the 
scores of a test case and then offset some number of them with stamps. 
  If I had the time to start analyzing, I would probably start with an 
even distribution of scores.


>> this is another reason for direct delivery on stamps.  Your stamp is an 
>> introducer.  It guarantees delivery to the inbox.  This is a win.  This 
>> mean customers don't have to worry about their mail getting through.
> 
> This is also a good reason for adaptive stamp sizes: You yourself can
> increase the chance for your mail to get delivered by just paying more.
> So on the one hand companies could accept small stamps in mails to their
> support-addresses to increase their chance that they miss no customer
> mail and on the other hand they could just pay enough for their own
> mailings to make sure they get read.

come up with a model, I would like to see it.  Personally I think a 
real-time dynamic pricing structure is far more appropriate because it can:

   o reduce stamp load on legitimate senders
   o increase costs on commercial and spam mail
   o make systems more resilient with regard to Moore's Law inflation
   o makes systems more resilient in the face of concentrated 
computation attacks (i.e. lots of zombies generating stamps aimed at a 
very small number of machines)

>> remember, transition costs are really expensive.  We want to do it good 
> 
> Using stamps as just another spam +/- indicator plus it being adaptive
> is IMHO a really simple transistion strategy.

read the archive.  There's lots of geek psychological resistance to 
using stamps.  I'm not going to go into it again.

---eric


PS, here are a few thoughts about the psychology of spam filters and 
their owners.  I need to repost it somewhere and I'm not sure where 
quite yet.

--- Spam filters are like dogs. ---

Spam filters are very much like dogs.  The similarity is apparent to 
anyone with experience with both.  They both need training, they both 
need daily care, and they both require dedicated owners.

But the similarity goes beyond this.  They are both used in competitive 
events where they are judged on how high a score they can get.  Spam 
filters are rated on the percentage accuracy.  One filter that I know 
quite well, CRM114 boasts a highly impressive 99.99% classification 
accuracy rate.  For dogs there are competitive obedience trials where 
they are stored on how accurately the performed exercise.  There are 
dogs I know scoring 198 out of 200 points at top-level difficulty dog 
obedience trials.

My dog on the other hand knows basic obedience, is reasonably well 
mannered yet barks her fool head off any time anyone makes a noise 
outside the house.  My spam filter accuracy runs around 90%. Fortunately 
I have the rest of the [link [url http://www.camram.org] camram system] 
to make up the difference and make spam a non-issue.

But back to the comparison.  What's the difference between my dog and 
the high-scoring obedience dogs?  Breed, temperament, and most 
importantly, owner dedication.  I'm willing to spend, in the beginning, 
the dedicated 15 week 5hrs/week effort it takes to train my dog to obey 
me.  In that same 15 weeks, the dog teaches me to hear a little bit 
about how it works.  Then, I spend the rest of its life communicating in 
the way that it wants to hear and reinforcing good behavior whenever 
possible.  As a result, I have a reasonably well-behaved dog that is not 
the win any prizes at an obedience trial.

In contrast, the high-scoring obedience folks work with their dogs four 
or five hours every single day, really intense training of both 
themselves and the dog so that they can get those high scores.  They get 
inside the dog's head and understand how it learns, how it will best 
hear the trainer.  This training process is the owner's life. they live 
with and for the dog.

How does this relate to spam filters?  High scoring spam filtration only 
happens if you dedicate your life to the spam filter, work with it every 
single day, and learn how to train it in the way it wants you to.   A 
spam filter is not something you can train intensively in the beginning 
and then just kind of leave alone.  It needs constant attention in order 
to keep it working right.

There's one more set of comparisons about spam filters and dogs.  Dogs 
and spam filters both have accidents and leave something unpleasant 
where you need to deal with it.  A major difference is that dogs can 
train you to recognize their signals and need to go outside so that they 
won't have accidents.  Spam filters keep giving you little presents in 
your mailbox every so often.  Dogs and spam filters also chew on things 
you don't want them to chew on.  With dogs, you usually know when 
they've chewed on something.  Not so with spam filters.

When it comes right down to it, most people find this concept of living 
for a piece of software repugnant.  They want to come into work, get the 
job done and that does not involve satisfying the attentional demands of 
spam filters.  Most people would also agree that any system which loses 
information silently or forces you to go through all the spam anyway to 
find what was lost is flawed.

given the choice, I, like most people, prefer to live with a dog because 
you get something worthwhile back from that relationship.