Page History

...

Social network talk:

• pretty useless spamfiltering-wise at least; not any spam orientation at all
  

...

• corpus analysis, from Hotmail's feedback loop
  • volunteers classify random samples of their mail as spam or good; tens of thousands of hand-classified messages per day; large "unbiased" (???) sample of spam
• additional analysis on two sets of spam:
  • about a year between the two
  • products sold, exploits used, trends
• viagra types: 17% 2003, to 34% 2004
• graphic porn down: 13% to 7%
• exploits: increasing rapidly, 1.33 exploits 2003 to 1.73 in 2004
• word obscuring: up to 20% in 2004
• URL chaffing, adding good URLs to spam: not there in 2003, 10% in 2004 – anti-SURBL attack
• Spammers are putting more work into each spam
  

Introducing the Enron Corpus:

• 1.3million messages originally; removed msgs with "integrity problems", replaced usernames etc
• http://www-2.cs.cmu.edu/~enron
• 200,399 useful, non-dupe messages
• 158 messages, 1,268 msgs/user
• missing message headers, so not much use for spam filtering; Exchange-mangled; no HTML. still, maybe good for "body" rules and FP avoidance
• no mention how much of the corpus was spam
  

Larry Lessig:

• extraordinary amount going to tech fixes; very little going to how the law could address it
• compares govt attention to "pirate radio" creating static for large commercial stations, vs the spam problem
• multiple types of regulators: the law, social norms, the market, and architecture (example: windows in lecture theatre are closed to enforce paying attention to speakers)
• the law also regulates the other three
• (that was the wrong talk! starts again!)
• 1. "regulation is always multiple modalities"
• 2. "interests will react"
• 3. "special interests defeat general interests"
• in the old days, we had norms to defeat spam; that failed
• using code to fix; so far that's failed
• "the market will fix the problem"; ISPs trying to be the spam-free email provider; that's also failed
• CAN-SPAM: totally failed – even displaced effective state legislation
• not any single modality alone can fix it
• regulation is a restriction, plus somebody to enforce it
• CAN-SPAM: wanted truthful headers
• opt-out doesn't provide any way for you to know if you've really been opted-out
• enforcement: state AGs, ISPs, federal - centralised; too big though. they have better things to do with their time than bust spammers
• solution: marries legal/architectural/market
• legal: has two parts: (1) labels ("ADV" in the subject line)
• (2) a bounty
• (q: SEXUALLY-EXPLICIT tag is a label, already massively flouted by spammers. other labels would be flouted just as much.)
• architecture: filter code then blocks mails with "ADV"
• market: spammers would then have to incentivise people to receive their mail by sending offers they want (yeah right
• enforcement: spam will only be sent if you can be paid, so "follow the money" – part of CAN-SPAM states "the business that benefits is responsible"
• market in enforcement: bounty hunters who identify label-less spam (ah). amateurs, not law enforcement, large population
• during CAN-SPAM development: labels were undesirable. Reason: "labels are too effective", because e.g. Amazon would have to have labelled their ads (because there was no distinction between opt-in and opt-out) and would be filtered
• fundamental problem: corruption due to vested interests lobbying (cf CAN-SPAM)
• sees difficulties in differentiating
• q: tracing spam to the business that benefits often involves getting forwarding addresses from e.g. a CGI script running on a server in the Ukraine. *needs* law-enforcement power to get that IMO. a: "yes, and law-enforcement power is available, and jurisdiction problems are easy" (not sure about that! at least for the non-LE bounty-hunter case)
• q: opt-in would have fixed it, like it has in Australia; but DMA keeps emasculating the laws into YOU-CAN-SPAM. a: agrees that there are multiple answers, but prefers not requiring opt-in across the board and uses the UCE definition as it allows political speech without adding to their costs. (I disagree, personally; the "UBE" definition works for me --jm)
• Jon Praed: enforcement requires tremendous resources, and in some cases you've got to get to that IP address within 7 days to get those logs, with LE power. This is not easy. Notes that spammer margins are incredibly low, and those bounties as a result would be small and/or hard to get.
• JP again: also suggests labels to label "good" commercial mail, personal mail, and then leave over "unknown" mail – which is then suspect. also suggests that the *headers* are the labelling, in reality.
• q: "special interests always seem to wipe out general interest on this issue in laws. what can we do, law-wise?" "my brand is pessimism", "there was this moment, when they passed CAN-SPAM, when legislators were keen to fix it – then the special interests came in".
• observation from audience: spots the parallel between UK and Pirate radio in the late 60's, which also passed a McCain anti-advertiser provision to deal with it.
• Dave Crocker: believes that the suggestion would result in little real effect on spammers, and quite a heavy hit on legit businesses
  
  

Hal Varian:

• "who is annoyed by spam?" one-fifth of US residents acknowledge buying products from spam; 77% considered spam an annoyance. That means 23-30% don't find it annoying! who are they?
• before federal DNC: multiple lists, state, DMA, company-specific
• federal DNC: lots more teeth; major fines, enforcement
• mapped DNC lists (with last 4 digits redacted, obtained via FOIA) against census data
• very popular in predom-asian areas
• income under $10k/yr: very low prob of signup
• income over $100k/yr: very high
• lots more interesting correlations, too many to write up
• income, education, number of kids are main significant determinators
• almost no corr between having internet and signing up for the DNC!
• est. signup rate for do-not-email: 31.5% (iirc)
• summary: telemktg, spam annoy same people
• 70% of variation in signups can be expl'd with only 4 vars: median income, presence of teens, education, presence of state list that was merged into fed DNC
• not many ppl used state lists, even though some were effective, and cheap. seemingly small costs can be big barriers
• DNC: effective because it had teeth ($11k fine), lots of publicity, and nationwide.
• q: "how do income and education correlate?" a: regression determines independent effects, accounts for correlation
• q: "do tmers target upper-middle-class?" a: yes, very much! q: "different to spam then, since spam is a lot more scattershot" a: agreed. also note that targeting sometimes aimed at *less* middle-class consumers, on the basis that they want to sucker people sometimes
• q on statistical reliability, answered, seems satisified
• q: about getting the NPA data via FOIA: apparently Telcordia asks for a fortune for that! also wondered if he'd checked in Europe. a: not yet, interested in the idea
• q: "what about using same techniques to find who benefits from spamming?" a: lots of interesting questions about looking at the spamming industry from an economic POV
  

Nicola Lugaresi: EU vs Spam - a legal response

• why legislate? social norms, market, self-regulation, self-help, have all failed; "code has failed"; law wants it chance to fail
• law probs: lack of jurisdiction and/or intl cooperation, lack of enforcement, lack of coord with other tools, bad laws
• EU law: 3 main goals: practical: fight spam; ethical: protect privacy, and state its relevance; political: don't just trail the US, lead sometimes!
• approach: optin and "soft opt-in" (transactional – in ctx of sale, similar products, same company; opp to object when email collected); no disguising identity; valid address to cease further comms
• other tools: labels, registries, spam boxes, codes of conduct – not convinced by any of these
• compared 2 antispam cases, one in Paris, a property case between spammer and AOL,MS; one in Napoli, a privacy case for an individual – E1000 damages and E750 costs paid to the individual!
• spam definition in EU: not bulk, just direct mktg not just "commercial". too narrow in my opinion
• not great against hard spam (proxy-abusers etc.): needs other approaches
• q: worried about definition of spam narrowed to direct mktg, and not bulk: a: agreed
• q: if I recall correctly not protecting corporate accounts, just individual "natural person" accounts at an ISP? a: not in Italian transposition at least, may be just in the Irish version (or my misreading
• q: role accounts considered? a: his opinion, yes, in Italy, that's the case. good news
  

No-Email-Collection flag: Matthew Prince, Unspam LLC

• Lessig-influenced presentation style – "law professors get paid based on the number of slides produced"
• CAN-SPAM removes individual rights to sue spammers, one thing that's really been effective so far. not good news
• proposes a new meta tag: no-email-collection
• cute cat photo!
• missing a huge chunk of the spam pipeline; we're focussing on the proxy-to-recipient part of the chain. focus on the address scraping!
• definition of spam: hard. but harvesting email addresses: everyone agrees that's a no-no

• Wiki Markup
  <meta name="no-email-collection" terms="\[url of terms page\] />

• http://www.unspam.com/noemailcollection
• project honeypot: like my script – email cookie addresses for scrapers
• create subdomains for honeypots and point them at Unspam's server! they'll collect the data
• generating a public corpus!
• code licensed under GPL
• http://www.projecthoneypot.org/ , http://www.unspam.com/
• q: "scraping through zombies?" a: yes, but that'll increase potential costs for spammers (hmm)
• q: "two classes of spammer lists: resold email addrs as well as scraped"? yes, but getting one class is useful
• q: "can a meta tag be enforceable? is a clickwrap license legally viable when it occurs between two computers?" a: if it's a community norm, that can improve legal viability; also CAN-SPAM specifically forbids scraping; also it may cause the spammer to think twice about this
  

Paula Bruening, CDT: Tech Responses to the Problem of Spam: Preserving Free Speech and Open Internet Values

• CAN-SPAM not entirely working
• worried about antispam tech hurting speech capabilities on the net
• concerned that "only popular speech will be delivered"
• what's key: tech must not be only part; devs must think of access issues; "let a thousand flowers bloom"
• q: "spam filters won't block websites" a: yes, but urgent updates do require email
• q: "is the CDT advocating political UBE?" a: good question, no answer
  

Barry Leiba, IBM: a multifaceted approach to spam filtering

• cowritten with Nathaniel Borenstein (woo!)
• "The Anti-Spam Gauntlet": describes exactly the SpamAssassin philosophy
• cooperation required with others
• open standards are required and are key to implementing anti-spam measures
• (I suggested open source as well as open stds
• Daniel: patents kill open standards; a from NSB: IBM are committed to "respecting IP rights" but not letting these stop open standards use by open source and other parties
  

SpamGuru:

• "enterprise-class anti-spam filter", but aren't we all
• centralized filter with personalized performance
• includes a "Bulk Mail Manager" for outbound *bulk* mail, interesting
• uses a "DNS analysis" step which sounds like it performs SPF checks
• DNS and domain analysis: check open relays, reverse DNS lookups and static IP tables; mail from dyn IPs; recency of dom registration; probabilistic analysis of Received trail
• bayes learning also feeds blacklist/whitelist; AWL is actually probabilistic
• "plagiarism detection": signature based really: "fast analysis of common k-grams"; learns from few examples; almost guaranteed not to be a FP; high FN rate though
• text classifier: Linear Discriminant: regularized linear classifier; approximates SVM
• Chung-Kwei (which rocks): really really effective: 86% with < 0.01% FPs on their test corpus
• test: corpus: 173k msgs, 130k spam, 42k good
• spam defn: UCE (not UBE). cleaned repeatedly
• combining algorithms: right with SpamAssassin dogma
• nice graph of aggregated performance; 96% with < 0.01% FPs
• SpamAssassin TODO: we need to add short-circuiting again!
• http://www.research.ibm.com/spam
• q: "what period, who were the 100 users?" a: users at IBM Watson
• q: how do you get your "recency of domain registration" data? a: straight from WHOIS
  

Richard Clayton: Stopping Spam by Extrusion Detection

• from demon.co.uk
• ISPs can spot smarthost load going up, and suspect that there's a spammer active
• insecure customers main problem for UK ISPs
• ISP's real problem: blacklisting of IP ranges and smarthosts; rapid action is req'd
• hard problem to solve: expensive to examine outgoing content; legal issues with blocking, and FP may cost you customers; volume is not good indicator!; "incorrect" sender domain doesn't indicate spam
• solution: spot delivery failure errors (due to user unknown, remote blocks) in smarthost logs
• heuristics: "too many" delivery failures (40/day sufficient); ignore "bounces" – have null <> return-path; ignore "mailing lists" (most dests work, few fail)
• when first turned on, was finding 40 infected customers *per day*!
• http://www.cl.cam.ac.uk/~rnc1/
• q: "direct-to-MX spam? trapping port 25?" a: no we don't do that and don't mind about that, as much as spammers using our smarthost and getting that blocklisted
• q: "sending outbound (or parts thereof) through SpamAssassin?" a: SpamAssassin is too expensive (in terms of load)
• q: "hair-trigger nature of listing?" a: it's not automatic. there's always a manual verification, and it's usually very obvious at that step
  

Resisting Spam Delivery through TCP damping:

• by default, TCP allows sender to control rate of flow; sender can achieve highest speed permitted by network
• TCP damping tries to reduce net efficiency at the receiver side; more time, more bandwidth, more CPU cycles
• low pain for recipients, high aggregated pain to spammers
• need to do this at TCP layer; higher and lower aren't useful
• even with tarproxy or similar, a smart spammer can blast the entire message to your TCP layer in one blat, even if you're tarpitting at the application layer
• damping: increase sending time (delaying TCP packets); consume network bandwidth (request more packets)
• increase delay: set adv_win = 0; fake congestion; delay outgoing ACKs (TCP conn terminates after 14 retries). cost at receiver: long idle TCP conn
• increase bandwidth costs: request more retrans.; request more ACKs – reuse sequence numbers, use seqs that won't be used in this conn; send packets in reverse order. cost: about 1:1 ratio
• used SpamAssassin at delivery time to estimate spamminess! mostly headers during early SMTP conversation, but you can use body rules before "250 Message Accepted for Delivery"
• q: economics. "increases senders costs, but not a transfer to the recipient." a: there are no existing techniques to do this, and TCP damping must work in existing system.
• q: if I was a spammer, and I figured out you were TCP damping, I'd ignore your advertised windows and blat entire message, hurting the network overall. a: sure, but hurting the spammer's bandwidth like this is worth it
• q: but this encourages broken TCP implementations. a: but a broken TCP stack still won't get their spam delivered
• q from John Levine: TurnTide does exactly this technique by narrowing the TCP window on the spammer's connections.
• q: why not just use delayed ACKs? a: because it's not entirely as effective as the other techniques
  

AOL hashing:

• I-Match: large corpus; lexicon generation
• intersection of document and lexicon gives signature
• trad I-Match lexicon generation: reject v frequent and hapaxes
• use "Mutual Information" as a measurement of fitness to avoid overlapping rules
• use multiple lexicons to avoid randomization from having an effect
• generate multiple lexicons, by removing random entries from an original lexicon
• also: distributional word clustering (Information Bottleneck) for lexicon selection (Terms with similar class distribution of P(spam|term))
• q: "'cluster' selection" – is that reports from live users? yep
• q: "FP rate?" a: very very low
  

Distributed, collaborative spam filtering:

• TCD, yay
• definition: "spam is email that the recipient is interested in receiving". we disagree, of course
• P2P approach
  

Reputation network analysis for mail filtering:

• 75% of semweb data is FOAF files
• using web of trust
• a bit like http://web-o-trust.org/ , but not yet workable with email addrs since there's no spoofing protection
  

On attacking statistical spam filters:

• spammers wanted to evade bayes
• tokenization/obfuscation: turn out to be good spamsigns
• should not have used SpamArchive spam, due to its lack of headers, in my opinion; headers improve spam recognition greatly
• (correction: SpamArchive spam now does include headers, I missed that change – so that's not a big deal. Also, from talking to one author post-talk, he noted that they omitted the hdrs since the spam and ham each came from a different corpus, therefore a different set of hosts. If not ignored, those tokens would have been very obvious clues for the classifier.)
• pretty similar to http://www.cs.dal.ca/research/techreports/2004/CS-2004-06.pdf