An Economic Response to Unsolicited Communication Thede Loder

© 2004 Loder, Van Alstyne & Wash, all rights reserved. 2

Data on the Spam Problem • Estimated yearly loss to business $10 billion (Ferris

Existing or Proposed Solutions • Technological: – Filtering: Rule based (static or dynamic), Naïve

Problems with Legislative Solutions • Defining spam – Whose definition? – One size fits

Problems with Pure Technology Solutions • Filtering (Rules, Collaborative, Bayesian): – – – False

Beat a Filter with Another Filter berkshire marriott wireless © 2004 John Graham-Cumming “How

Modeling the Message Value • Each message has a party-dependent value: – value to

Terminology & Assumptions • [r, r] and [s, s] are bounds on reader and

Modeling Email Value r s r cs = cost to send cr = cost

Modeling Email Value r s r Incentive Compatible for Sender cs = cost to

Social Welfare Contribution r Welfare (W) = RS + SS cr s Received email

Interpreting Existing Solutions r Solution Sender cost Social cost Flat Tax Per email tax

Interpreting Filtering r Good mail blocked Bad mail passed : 0 < < 1

Perfect Filter r loss Wanted cr Unsent gain cs Filtered cs/ Definition A technological

Consider an “Attention Bond” Initially, the sender knows more about the message than the

Recipient Payoff with Bond r For any particular distribution, the recipient can remove the

Recipient Payoff with Bond r When ( b = p ) > 0, sender

So why not just seize the bond all the time? © 2004 Loder, Van

Comparison of Bond to Perfect Filter • • G distribution expected to be mostly

Bond permits discriminatory costs • More costly for spammers due to being collected more

Distribution G r “Good” G cr s W+ cs © 2004 Loder, Van Alstyne

Distribution B r cr B s “Bad” W+ cs © 2004 Loder, Van Alstyne

G Sent with Attention Bond r Unsent G G cr s For each distribution,

B Sent with Attention Bond r cr Unsent B B For a mostly unwanted

Welfare Basis - Attention Bond r G cr B cs + pb cs +

Welfare Basis - Perfect Filter r G B Sent & Filtered cr s Sent

The ABM in Action $$ A A N A M C MC MM Sender

Architecture for ABM © 2004 Loder, Van Alstyne & Wash, all rights reserved. 45

Additional Social Benefits • Direct marketing – Cheaper than traditional mail channels – Receiver

Conclusion • Screening action of bond forces senders to reveal their intentions • Creates

Slides: 43

Download presentation

Data on the Spam Problem • Estimated yearly loss to business $10 billion (Ferris Rsch) • ISPs estimate the cost of spam at $2 -$3 per user per month (IDG) • The Federal ‘Can Spam’ Act passed with near unanimous support; superceding 8 State laws, 6+ spam laws still pending in congress (NYT) • 55% of all email is now spam (Brightmail) • 29% of Americans report curtailing e-mail use (Pew Internet report) Spammer Alan Ralsky vows to carry on (NYT) © 2004 Loder, Van Alstyne & Wash, all rights reserved. 3

Existing or Proposed Solutions • Technological: – Filtering: Rule based (static or dynamic), Naïve Bayesian Filters, collective/community classification – Challenge Response: quasi-Turing tests, return address testing • Legislative/Regulatory: – Banning, labeling – Taxation, stamps • Economic – Computational challenge (proof of work) – Bond © 2004 Loder, Van Alstyne & Wash, all rights reserved. 4

Problems with Legislative Solutions • Defining spam – Whose definition? – One size fits all • Regulating: Banning and labeling – Enforceability, jurisdiction? – Costly to police and adjudicate – Labeling lacks incentive compatibility • Taxing: e-stamps – Blocks wanted along with unwanted email – Blunts cost-effectiveness of email as medium © 2004 Loder, Van Alstyne & Wash, all rights reserved. 5

Problems with Pure Technology Solutions • Filtering (Rules, Collaborative, Bayesian): – – – False positives, false negatives Increases spam email traffic Costly arms-race Consensus definition Blocks automated email • Challenge-Response (Reverse Turing Tests) – Cheaply hire real people – Unrecoverable costs © 2004 Loder, Van Alstyne & Wash, all rights reserved. 6

Modeling the Message Value • Each message has a party-dependent value: – value to the sender s – value to the recipient r • Range of s and r bounded by r, r, s, s Value to Sender (s) - + s 0 s Value to Recipient (r) - + r 0 © 2004 Loder, Van Alstyne & Wash, all rights reserved. r 8

Terminology & Assumptions • [r, r] and [s, s] are bounds on reader and sender values • Sender knows own value s and cost cs before sending, won’t send when s < cs • Recipient only knows her r value after reading and incurring cost cr • Filtering results in a fraction of messages not getting through © 2004 Loder, Van Alstyne & Wash, all rights reserved. 9

Modeling Email Value r s r Incentive Compatible for Sender cs = cost to send cr = cost to receive cr s • message is sent when s > cs • sender expects positive payoff r s cs region of positive probability © 2004 Loder, Van Alstyne & Wash, all rights reserved. 11

Modeling Email Value r s r cs = cost to send cr = cost to receive Wanted cr s Examples: - project update from business associate - long lost high school buddy says hi - friend’s wedding announcement - shipping update from ecommerce vendor - recruiting lead from a friend r s cs region of positive probability © 2004 Loder, Van Alstyne & Wash, all rights reserved. 12

Modeling Email Value r s r cs = cost to send cr = cost to receive cr s Unwanted Examples: - offensive pornography - message with attached virus - phishing, Nigerian scam r s cs region of positive probability © 2004 Loder, Van Alstyne & Wash, all rights reserved. 13

Modeling Email Value r s r cs = cost to send cr = cost to receive cr Unsent s • email unsent when s < cs • sender expects negative payoff r s cs region of positive probability © 2004 Loder, Van Alstyne & Wash, all rights reserved. 14

Modeling Email Value r s r cs = cost to send cr = cost to receive Wanted but Not Sent cr s Examples: - personalized loan application - customized news or subscription content - sales leads r s cs region of positive probability © 2004 Loder, Van Alstyne & Wash, all rights reserved. 15

Modeling Email Value r s r cs = cost to send cr = cost to receive cr s Unwanted and Not Sent Examples: - embarrassing news sent to wrong person - outdated or incorrect facts - meaningless details - accidentally telling your boss/advisor what you really think r s cs region of positive probability © 2004 Loder, Van Alstyne & Wash, all rights reserved. 16

Social Welfare Contribution r Welfare (W) = RS + SS cr s Received email northeast of the diagonal line makes a positive contribution to social welfare positive contribution to W negative contribution to W W+ cs © 2004 Loder, Van Alstyne & Wash, all rights reserved. 19

Interpreting Existing Solutions r Solution Sender cost Social cost Flat Tax Per email tax Tax administration E-stamps Per email stamp Stamp creation and collection Computational Challenge Compute cycles Fixed & marginal sunk cost Challenge Response or quasi-Turing Human time, compute time Fixed & marginal sunk cost Criminalizing, Do-Not-Spam list Evasion cost Enforcement Cost loss Wanted cr Unsent s gain cs Unwanted cs+ extra cost W+ © 2004 Loder, Van Alstyne & Wash, all rights reserved. 20

Interpreting Filtering r Good mail blocked Bad mail passed : 0 < < 1 email gets through loss cr Unsent s gain cs Filtered Waste cs/ W+ Sender decision: send when E(gross return) – cost > 0 No filter E(gross return) is s, cost is cs send when s – cs > 0 or s > cs With filter E(gross return) is s, cost is cs send when s – cs > 0 or s > cs/ © 2004 Loder, Van Alstyne & Wash, all rights reserved. 21

Perfect Filter r loss Wanted cr Unsent gain cs Filtered cs/ Definition A technological filter that: § Operates without cost § Makes no mistakes (no false positives, no false negatives) § Intuits and internalizes all reader preferences § Eliminates, prior to receipt, any email where r < cr Not all filtered email is waste! W+ © 2004 Loder, Van Alstyne & Wash, all rights reserved. 22

Consider an “Attention Bond” Initially, the sender knows more about the message than the receiver, so force them to reveal that private knowledge: • • Simple screening mechanism applied to unrecognized senders. Challenge demands an escrowed bond fee of amount . Recipient has sole discretion to claim or return , with expected bond forfeiture b=p. All proceeds b go to recipient. Effects: 1. A recipient-controlled variable ‘tax’ on senders, based on sender behavior and message content 2. Shift task from ex ante classification (hard) to ex post verification (easy). © 2004 Loder, Van Alstyne & Wash, all rights reserved. 23

Recipient Payoff with Bond r For any particular distribution, the recipient can remove the sender incentive to send email. The sender sends when s > cs+p. Any lesser value of s yields a negative expected payoff. Wanted cr s Unsent Unwanted Positive payoff to recipient Negative payoff to recipient W+ cs © 2004 Loder, Van Alstyne & Wash, all rights reserved. 26

Recipient Payoff with Bond r When ( b = p ) > 0, sender costs are cs+p. Increasing p eliminates incentive to send, blocking more of the email in the distribution. Wanted cr -p s Unsent Positive payoff to recipient Unwanted Negative payoff to recipient Now transfer p where p >0, > 0 cs+p © 2004 Loder, Van Alstyne & Wash, all rights reserved. W+ 27

Recipient Payoff with Bond r When ( b = p ) > 0, sender costs are cs+p. Increasing p eliminates incentive to send, blocking more of the email in the distribution. p increasing… Wanted cr -p Unsent Positive payoff to recipient Unwanted Negative payoff to recipient cs+p © 2004 Loder, Van Alstyne & Wash, all rights reserved. W+ 28

Recipient Payoff with Bond r When ( b = p ) > 0, sender costs are cs+p. Increasing p eliminates incentive to send, blocking more of the email in the distribution. p increasing… Wanted scr -p Unsent Positive payoff to recipient Unwanted Negative payoff to recipient cs+p © 2004 Loder, Van Alstyne & Wash, all rights reserved. W+ 29

Recipient Payoff with Bond r When ( b = p ) > 0, sender costs are cs+p. Increasing p eliminates incentive to send, blocking more of the email in the distribution. p increasing… Wanted s Unsent cr -p Positive payoff to recipient Unwanted Negative payoff to recipient cs+p © 2004 Loder, Van Alstyne & Wash, all rights reserved. W+ 30

Recipient Payoff with Bond r When ( b = p ) > 0, sender costs are cs+p. Increasing p eliminates incentive to send, blocking more of the email in the distribution. Wanted s Unsent cr -p Positive payoff to recipient Negative payoff to recipient Unwanted cs+p © 2004 Loder, Van Alstyne & Wash, all rights reserved. W+ 31

Recipient Payoff with Bond r When ( b = p ) > 0, sender costs are cs+p. Increasing p eliminates incentive to send, blocking more of the email in the distribution. Unsent Wanted cr -p Positive payoff to recipient Negative payoff to recipient Unwanted cs+p © 2004 Loder, Van Alstyne & Wash, all rights reserved. s W+ 33

Recipient Payoff with Bond r When ( b = p ) > 0, sender costs are cs+p. Increasing p eliminates incentive to send, blocking more of the email in the distribution. Unsent Wanted Positive payoff to recipient Negative payoff to recipient cr -p Unwanted cs+p © 2004 Loder, Van Alstyne & Wash, all rights reserved. s W+ 34

Comparison of Bond to Perfect Filter • • G distribution expected to be mostly ‘Good’ B distribution expected to be mostly ‘Bad’ Sender does not know r Sender knows to which distribution (G or B) his email belongs • Bond size is the same for both distributions © 2004 Loder, Van Alstyne & Wash, all rights reserved. 36

G Sent with Attention Bond r Unsent G G cr s For each distribution, the recipient can choose a policy with a seize probability px W+ cs cs + pg © 2004 Loder, Van Alstyne & Wash, all rights reserved. 40

B Sent with Attention Bond r cr Unsent B B For a mostly unwanted email distribution, the recipient is best off if they seize with a high probability (expected bond size b = p is large) s W+ cs cs + pb © 2004 Loder, Van Alstyne & Wash, all rights reserved. 41

The ABM in Action $$ A A N A M C MC MM Sender initiates server attempts amail message delivery Recipient’s Sender or. The …and the Sender Recipient not challenge mail Escrow Sender Mail onmail on server The whitelist, Whitelist, Server issues claims is issues bond delivered issues notification delivers bond-challenge bond posted bond to or authorization delivered original the authorization lets of Sender… it bond expire. issued message receipt The bond isis returned 44 © 2004 Loder, Van Alstyne & Wash, all rights reserved.

Additional Social Benefits • Direct marketing – Cheaper than traditional mail channels – Receiver gets the benefit, not printer and mailer – Costs less than snail-mail • • • Recipients have reason to publish contact information, not hide it Shifts arms race to crypto. Permits communication otherwise filtered (e. g. political speech) Tailors to an individual’s unique preferences. Reclaims the value of your address for you rather than the marketers. Helps you get off any list you don’t want to be on. © 2004 Loder, Van Alstyne & Wash, all rights reserved. 46

Conclusion • Screening action of bond forces senders to reveal their intentions • Creates possibility of greater trade volume, with both parties wanting to participate • Facilitates wealth transfers benefits readers more than unilateral veto. Can beat a Perfect Filter. • Shifts the arms-race to crypto • Individualized, works with existing uses • Returns control of the mailbox to its owner © 2004 Loder, Van Alstyne & Wash, all rights reserved. 47