New Kind of Spam 'Un-Training' Filters?

Zaphod2016 writes to tell us the Wall Street Journal is reporting that email in-boxes are under a new kind of spam attack. This new spam has confused many people due to its lack of advertising, viruses, or request for personal information. One popular theory is that these innocuous blocks of text, often drawn from popular literature, are being used to "un-train" spam filters to allow more malicious spam through in the future.

Comments

Un-training? Hardly.

[pclminion]

Bayesian and other filters do not rely on "spammy" words alone -- they also rely on "unspammy" words, and spammers have no idea what those words are because each person receives different email.

A scenario, with made up (but plausible) numbers: Suppose you're a developer of a Linux driver for the Bozodrive 1000. The majority of your legitimate email comes from Linux driver development mailing lists. A full 50% of those emails contain the word "IRQ." 99% of the emails contain the word "driver," and 15% contain the word "Johannsen" which is in the signature of one of your friends. And precisely 0% of the emails containing any of these terms have ever been found to be spam.

Any decent spam filter will give a huge weight to the presence of these "unspammy" words, because of the extremely high probability of emails containing them to be non-spam. The presence of randomly selected confusion words in empty spams is not going to affect these frequency counts.

In order to defeat a filter by confusing it, the spammer must guess what the SPECIFIC non-spam words for that PARTICULAR email user are, and then produce bogus, spam messages containing those words in the appropriate frequencies. This will cause the classification counts for those words to become more equalized, and the value of those words in determining spammyness to be greatly reduced. However, this is an impossible task unless the spammer has access to the actual emails of the target.

Perhaps the intent of the empty spams is to confuse the filters, but whoever devised the method has no understanding of how these things actually work, whatsoever.

Re:Not very effective and may be easy to work arou

[pclminion]

By having a baysian filter forget over time, it also helps shrink down the database and helps it adapt as the contents of spam change over time.

Having the filter forget is the ONLY effective policy. In statistical filtering, it is certainly NOT true that more data == better results. You want a sample of data that most accurately represents the sort of content you are receiving RIGHT NOW. I completely purge my Firefox Bayesian database every couple of months and retrain on recent emails only. The result is ALWAYS an increase in accuracy, particularly a reduction in false positives.

No, unless people send that text to you.

[khasim]

I still flag crap like this as spam, so it seems like it'd train my spam filter to have more false positives, no?

No. Unless the people you usually corresponde with also include blocks of the same text.

The only way to increase the false positives is to get the spam filter to learn the words that usually appear in your legitimate messages.

Since the spammers have no way of knowing what those words are, there is no way they can bypass your filters ... and still be effective in getting through any one else's filters.

Re:Other way around?

[TubeSteak]

My limited experience is that whatever filtering Hotmail uses has been allowing lots of Spam to slip through in the last few weeks.

Anyone else?
How's Yahoo & G-Mail been doing?

Re:I just thought they were weird.

[CohibaVancouver]

be interested to know how many people put up money for products / services they were spammed with.

Quite a few, apparently.

I read one article which claimed that one spammer in particular "received 10,000 credit card orders in one month [snip] each for $39.95 US."

So that's nearly $400,000 per month. Nice work if you can get it.

Source:

http://www.cbc.ca/story/business/national/2005/04/ 08/spam-050408.html

Re:The text comes from the Gutenberg Project

[Ed Avis]

If the spammers are now sending round Gutenberg texts, this is entirely appropriate. Project Gutenberg caused probably the first ever spam, when Michael Hart launched the project by trying to mail everyone on ARPANET with the U.S. Declaration of Independence. (source)

Re:The text comes from the Gutenberg Project

[letxa2000]

think that is the point. They want to either poison those words so you get more false positives or they want to push other REAL spam related words out of the "this is spam" dictionaries. Maybe both. If these messages had some common theme, they would all get blocked and would have no net effect. They need you to click "this is spam" to poison your filters. Question is, does it work?

Answer is: No, it won't. At least not with Bayesian. The only way to mess up a Bayesian filter is if they can send you messages that are heavy in words/terms that often appear in your good email. And that's going to vary from user to user. Unless you're sending me the exact words that I use in my daily emails, adding a plethora of other words is not going to make my filter any less accurate or create more false positives. It will either let my filter recognize your "poison" as spam itself or, at worst, be neutral.

My Bayesian filter, among other things, considers an excessive number of infrequently/never used terms as a characteristic that is itself subject to Bayesian classification. So while the "poison words" have no statistical effect on my filter, the fact that a bunch of unusual words are found in a message is going to increase the chance that my filter correctly recognize the message as spam.

My spam was constantly growing through about December of last year. This year, it seems to have leveled off. Sure, I'm still getting just under 20,000 per month which sucks, but I see almost none of them and according to my spam stats, the spam has leveled off. Hopefully this is the plateau before it falls. :)

I still want to know: Who are the idiots who BUY spammed products???

Re:Other way around?

[badasscat]

How's Yahoo & G-Mail been doing?

Here are actual samples of emails that Gmail and Yahoo have let through to my inbox over the past couple days. First, Gmail:

Wells, who has had a rather similar historyand who obviously owes something to Dickens as novelist. In some ways his outlook is verysimilar to Dickenss. No one who is really involved in the landscape ever sees thelandscape. To Chesterton the poor means small shopkeepers andservants. There is nothing psychologically false in this, either. No one who is really involved in the landscape ever sees thelandscape. It is easy to imagine what the young woman would have said to this inreal life. And given the FACT ofservitude, the feudal relationship is the only tolerable one. Theother point is that Dickenss early experiences have given him a horrorof proletarian roughness. They, and the men, always spoke of me as the younggentleman. It is one of the stockjokes of English literature, from Malvolio onwards. Buthe is remarkably free from the idiocy of regarding nations asindividuals. So were all the characteristic English novelists of thenineteenth century. The last thing anyone ever remembers about the books is theircentral story. Nevertheless hislist of most hated types is like enough to Wellss for the similarity tobe striking. A change of heart is in fact THE alibi of peoplewho do not wish to endanger the STATUS QUO. There is nothing psychologically false in this, either. Pickwick and the servant should be Sam Weller. It is noticeable thatDickens hardly writes of war, even to denounce it. Therewere no labour-saving devices, and there was huge inequality of wealth. In Dickenss novels anything in the nature of work happens off-stage. And, on the whole, his attacks on good society are ratherperfunctory. But byorigins and upbringing Thackeray happens to be somewhat nearer to theclass he is satirizing. Here perhaps Gissing is influenced by his own love of classical learning. In a rather different sense his attitude to life is extremely unphysical. It is usual to claim him as a popularwriter, a champion of the oppressed masses. Dickens would be quite incapable of this. Compare any lawsuit in Dickens with the lawsuit inORLEY FARM, for instance. I do consider the young ooman, sir, said Sam. Here the contrast between Dickens and, say, Trollopeis startling. It is true that not all his novelsare alike in this. He getshimself arrested in order to follow Mr. Progressis not an illusion, it happens, but it is slow and invariablydisappointing. If his palms are hard from work, they let him in; if his palms aresoft, out he goes. It is perhaps more significant that he shows noprejudice against Jews. At first sight this statement looks flatly untrueand it needs some qualification. A modern manservant would neverthink of doing either. There arepractically no friendly pictures of the landowning class, for instance. If one wants a modern equivalent,the nearest would be H.

Attached to the above was an image file that contained an obvious ad. So to Gmail, this apparently looks like a regular text email that happens to have an attached image.

(You can argue about how effective this is, since Gmail thumbnails all images, meaning you'd need to click a separate link to open it and read it.)

Now Yahoo, where I get approximately 1,000 messages to my bulk folder per day - this is the only one that's gotten through to my inbox in the last day:

FROM THE DESK OF Mrs Queen Adams
BANK OF AFRICA [BOA]
OUAGADOUGOU, BURKINA FASO.

DEAR FRIEND,

I AM HOPEFUL THAT THIS MAIL WILL REACH YOU IN GOOD CONDITION OF
HEALTH.I AM MRS QUEEN ADAMS A STAFF OF BANK OF AFRICA AND A BURKINABE RESIDENT
IN BURKINA FASO ALSO.IN THE BANK WHERE I WORK AS AN AUDITOR,I
DISCOVERED AN ABANDONED SUM OF MONEY AMOUNTING TO 15.2MILLION DOLLARS BELONGING
TO DR GEORGE BRUMLEY WHO UNFORTUNATELY DIED IN THE PLANE CRASH OF UNION
TRANSPORT AFRICAN FLIGHT BOEING 727 IN KENYA, EAST AFRICA ON SUNDAY

Re:Spam is dying

[dodobh]

I work for a fairly large email service provider. Spam isn't dying by any means. We just doubled production hardware last week to have enough smtp listener processes to be able to accept email. Bayesian is nice for the single user. For an ISP, it isn't. ISPs are bearing the brunt of the expense right now. The day I fear is when ISPs start to go under, or start charging for spam filtering, or simply stop.

Those boxes are running at sustained loads of 40+ and are CPU bound. That's a bit rare in the email world, as you would know if you have ever run a non trivial system in production.

The spammers will send more spam is something that we have been observing in reality. I have seen AOLs numbers, and they are merely two orders of magnitude bigger than ours at the moment.