An Economic Solution to Spam

Document Sample
An Economic Solution to Spam Powered By Docstoc
					An Economic Solution to Spam

  Thede Loder, Marshall Van Alstyne,
             Rick Wash
Spam Examples
Want these guys in your mailbox?




                                       Spam King: Alan Ralsky
                                       spewed tens of thousands of
                                       e-mail sales pitches per hour,
                                       bringing on the wrath of
                                       Verizon. Tech. Review Aug ‘03



  Richard Colbert, spammer. NYT 09-28-03
          The Spam Problem
• Estimated yearly loss to businesses $10 billion
  (Ferris Research)
• ISPs estimate the cost of spam at $2-$3 per user
  per month (IDG)
• 6+ spam laws are pending in congress (NYTimes)
• 8 states have already made spam laws including
  recent CA law (NYTimes)
• 50% of all email is now spam (Brightmail)
  Existing or Proposed Solutions
• Two Categories: legislative, technological
• Technological Candidates:
   – Filtering: Rule based (static or dynamic), Bayesian
     Filters, collective/community classification
   – Challenge Response: quasi-Turing tests, return address
     testing, computational challenge (solve hard problem)
• Legislative Candidates:
   – Banning, labeling
   – Taxation, stamps
 Problems with Pure Technology
           Solutions
• Filtering:
   –   False positives, false negatives
   –   Costly arms-race
   –   Consensus definition
   –   Shuts down exchange
• Challenge response:
   – Requires human interaction
   – Costly arms race
   – Blocks automated emails
     Problems with Legislative
            Solutions
• Banning and labeling:
  – Enforceability (jurisdiction?)
  – Costly to police and adjudicate
  – Labeling lacks incentive compatibility
• Taxation and Stamps:
  – Blocks wanted along with unwanted email
  – Kills cost-effectiveness of email as medium
…and Spam is Getting Worse


     What‟s going on?
                              Modeling Email
• Each email has a party-dependent expected value:
      – value to the sender s
      – value to the recipient r
• s and r bounded by Vs and Vr (upper and lower)
Value to Sender (s)
       -                                              +
                                            __
                         Vs
                         __             0   Vs
Value to Recipient (r)

       -                                              +
                                                 __
                         Vr
                         __         0            Vr
           Basic Assumptions
• r and s are expected values
• Sender knows s before sending
• Sender does not know r
• Sender knows or learns their marginal cost, and
  will not send when s < cost
• Base cost to send cs is the same for all
  distributions
• Recipient knows r upon receipt but only after
  losing an unavoidable receiving/processing cost cr
 Email Value - Single Distribution
          r (value to recipient)   __
                                   Vs
                                        __
                                        Vr          cs = Cost for sender to send
                                                    cr = Cost for recipient to
                                                              receive (sender must
                                             cr               send)
                                             s
                                        (value to sender)



                                                            = region of positive
__                                                          probability in the
Vr                                                          distribution for email with
     Vs
     __                                                     values (r, s)

            cs
       Derivation of Payoffs 1
• General form: distribution of point probabilities

       f ( s, r )   such that    f (s, r )  1
                                A

• We assume f is uniform in each dimension
  and independent of s and r.
        f ( s, r ) = distribution specific constant (k)
       Derivation of Payoffs 2
• Baseline Recipient Surplus is
          RS0   (r  cr ) f (s, r )dsdr
                  A

                  k  (r  cr )dsdr
                      A

                            1       1
         Where        k        
                         Vs  Vs Vr  Vr
      Derivation of Payoffs 3
• Similarly, we can determine the Sender
  Surplus (SS) and the Social Welfare (SW)

            SS0  k  ( s  cs )dsdr
                     A


               SW0  RS 0  SS0
Social Welfare Contribution
    r
                   For a single email:

                      Social Welfare (SW) = RS + SS


              cr      Sent email with (r, s) northeast of the
                      diagonal line makes a positive
              s       contribution to social welfare if
                      received


                             = positive contribution to SW

              SW+            = negative contribution to SW


        cs
Interpreting Existing Solutions
Flat Tax, Computational Challenge      Filtering (all types)
            r                                r



                 loss                             loss

                                                                         cr
   Unsent                           Unsent
                                                                         s

                         Unwanted                            Filtered
                 loss                             loss        Waste



                     cs+ tax                          cs/n              SW+
                cs                               cs
                    Existing Solutions 2
Challenge Response - quasi-Turing Test     Criminalizing Spam
                r                                 r



                     loss                              loss


       Unsent                            Unsent



                              Unwanted                          Unwanted
                     loss                              loss



                         cs+ test cost                     cs+ risk of penalty
                    cs                                cs
                            Perfect Filter
                                              r

Definition: A technological filter
     which:
1) Operates without cost                                Received
2) Makes no mistakes (no false
     positives or false negatives)                                  cr
3) Intuits and internalizes all reader   Unsent
     preferences                                                    s
4) Works for an arbitrarily large
     number of different distributions                  Filtered
5) Eliminates, prior to receipt, any
     email where r < cr

                                                                   SW+


                                                  cs cs/
“In terms of individual and aggregate
social welfare, a system that facilitates
valuable exchange will generally dominate
a system that grants unilateral veto power
to either party”
Not all filtered email is waste
           r

                                     = positive contribution to welfare

                                     = negative contribution to welfare

                               cr    = not sent (s < cs)
  Unsent
                               s     = unwanted by recipient (r < cr)

                    Unwanted
                                      Some spam has positive
                                          contribution

                               SW+



               cs
If it‟s not waste, don‟t destroy it
             r

                                       • filter nothing
                                       • set bond = 0

                                 cr
    Unsent                             • filter everything
                                       • set bond to transfer sender surplus!
                                 s
                      Unwanted
                                       • filter everything
                                       • set bond > Vs
                                       • (stop using email entirely)
                                 SW+



                 cs
 The Attention Bond Mechanism
• Follows from mechanism design - an economic, rather
  than pure technological or legislative design
• Simple screening mechanism
• Challenge demands an escrowed monetary bond of the
  amount  (a „fee‟)
• Recipient has sole discretion to keep or return 
• Effect: a recipient-controlled variable „tax‟ on senders, by
  type.
• „Tax‟ proceeds go to recipient
• Used for non-whitelisted senders
Architecture for ABM
                    Recipient Payoff
                                              r

For any particular distribution,
the recipient can remove the
sender incentive to send emails
for which s < cs+p, while at the                       Wanted
same time gaining p.                                             cr
                                     Unsent
           Here, p = 0                                            s

      Positive payoff to recipient                     Unwanted

      Negative payoff to recipient
                                                                  SW+



                                                  cs
                    Recipient Payoff
                                              r

For any particular distribution,
the recipient can remove the
sender incentive to send emails
for which s < cs+p, while at the                          Wanted
same time gaining p.                                                cr
                                     Unsent                          cr -p
                                                                     s

      Positive payoff to recipient                        Unwanted

      Negative payoff to recipient
                                                                     SW+

                                                       cs+p
                                                  cs
                    Recipient Payoff
                                          r

For any particular distribution,
the recipient can remove the
sender incentive to send emails
for which s < cs+p, while at the                          Wanted
same time gaining p.                                                   cr
                                     Unsent                             cr -p
                                                                    s

      Positive payoff to recipient                     Unwanted
      Negative payoff to recipient
                                                                    SW+

                                                   cs+p
                                              cs
                    Recipient Payoff
                                         r

For any particular distribution,
the recipient can remove the
sender incentive to send emails
for which s < cs+p, while at the
                                                       Wanted
same time gaining p.                                            cr
                                     Unsent                      s
                                                                 cr -p

      Positive payoff to recipient
                                                      Unwanted
      Negative payoff to recipient
                                                                 SW+

                                                   cs+p
                                              cs
                    Recipient Payoff
                                        r

For any particular distribution,
the recipient can remove the
sender incentive to send emails
for which s < cs+p, while at the
same time gaining p.                               Wanted    cr
                                     Unsent
                                                              s
                                                              cr -p

      Positive payoff to recipient
                                                   Unwanted
      Negative payoff to recipient
                                                              SW+

                                                 cs+p
                                            cs
                    Recipient Payoff
                                       r

For any particular distribution,
the recipient can remove the
sender incentive to send emails
for which s < cs+p, while at the
same time gaining p.                              Wanted    cr
                                     Unsent
                                                             s
                                                             cr -p
      Positive payoff to recipient
                                                  Unwanted
      Negative payoff to recipient
                                                             SW+

                                                cs+p
                                           cs
                    Recipient Payoff
                                       r

For any particular distribution,
the recipient can remove the
sender incentive to send emails
for which s < cs+p, while at the
same time gaining p.                                       cr
                                                  Wanted
                                     Unsent
                                                            s

                                                            cr -p
      Positive payoff to recipient

      Negative payoff to recipient               Unwanted
                                                            SW+

                                                cs+p
                                           cs
                    Recipient Payoff
                                      r

For any particular distribution,
the recipient can remove the
sender incentive to send emails
for which s < cs+p, while at the
same time gaining p.                                     cr
                                                Wanted
                                     Unsent
                                                          s

      Positive payoff to recipient                        cr -p

      Negative payoff to recipient             Unwanted
                                                          SW+

                                               cs+p
                                          cs
                    Recipient Payoff
                                     r

For any particular distribution,
the recipient can remove the
sender incentive to send emails
for which s < cs+p, while at the
same time gaining p.                                     cr
                                     Unsent     Wanted
                                                          s

      Positive payoff to recipient
                                                          cr -p
      Negative payoff to recipient             Unwanted
                                                          SW+

                                              cs+p
                                         cs
Extended Model and Comparison
• How do RS, SS, and SW compare?
• Multiple Distributions
• Perfect Filter – average cost per receipt goes
  to cs/
• ABM - Cost to sender increases by pd .
  Cost now cs+ pd 
   The no intervention baseline

                                      vr v s
Reader Surplus defined:   RS0  k            (r  cr )drds
                                    v r cs




                           RS 0 
                                  vs  cs   vr  v r             
                                                                cr 
Baseline surplus                              
                                  vs  v s   2                   
  The Reader‟s choice of screen
                                       vr vs
Reader Surplus defined:      RSB                (r  cr  p )drds
                                     v r cs  p

                                   1                vr  v r      
Optimal Screen:                
                               
                                      vs  cs   
                                                               cr  
                                                                      
                                  2p                   2          
                                   1                vr  v r      
Optimal Seize Rate Policy:
                               
                              p      vs  cs   
                                                               cr  
                                                                      
                                  2                   2          
                                                                            2
                                                      vr  v r      
                                        vs  cs   
                              1
Reader Surplus:     RSB                                         cr  
                                                                        
                          4vs  v s                    2          
Always:                      RS B  RS 0
                   The Perfect Filter
                                         vr vs
Reader Surplus defined:       RSPF              (r  cr )drds
                                        cr cs /




                                                         cs 
                                       vr  cr   vs  
                                                   
                                                    2
Reader Surplus:
                                                           
                              RSPF   
                                        2vr  v r vs  v s 



                                                       vr  v r      
Bonding wins if:    RS B  RS PF (2 p )  8vs  cs 
                                          2
                                                                  cr 
                                                       2             
 Social Planner‟s choice of screen
                                           vr vs
Welfare defined:   WSP  SSSP  RSSP              (r  cr  s  cs )drds
                                       v r cs  p




                                     1  vr  v r      
Optimal Screen:                     
                                   *
                                                   cr 
                                     p 2              


                                       3
              Always:              WB  WSP
                                       4
      Comparison of ABM to PF
        2-Distribution Case
•   V distribution expected to be mostly „Valuable‟
•   W distribution expected to be mostly „Waste‟
•   Sender does not know r
•   Sender knows to which distribution (V or W) their
    email belongs
Distribution V
      r


                   “Valuable”


               V
                   cr

                   s




                   SW+



          cs
Distribution W
      r




                   cr

                   s
               W
                   “Waste”



                   SW+



          cs
       V Sent With Attention Bond
                                        r




                                       Unsent
                                                     V
                                         V
                                                             cr

                                                             s


For each distribution, the recipient
can choose a policy with a seize
probability px
                                                             SW+



                                            cs   cs + pv 
       W Sent with Attention Bond
                                      r




                                                            cr

                                                            s
                                          Unsent W      W
For a mostly unwanted email
distribution, the recipient is best
off if they seize with a high
probability (greater p product)
                                                            SW+


                                          cs     cs + pw
Recipient Surplus - ABM
          r



                          V
                                    W
                                        cr

                                        s
                                    W




                        cs + pw        SW+
                   cs + pv 
              cs
Recipient Surplus – Perfect Filter
                r



                                  V

                                       W   cr

                                           s



                                  cs /w
                                           SW+
                         cs /v

                    cs
Welfare Basis - ABM
        r




                             V
                                     cr

                                     s
                                 W




                         cs + pw    SW+
                 cs + pv 
            cs
Welfare Basis - Perfect Filter
              r



                                V

                                      W        cr
                         Filtered Waste
                                               s
                                    Filtered
                                     Waste



                                cs /w         SW+
                       cs /v

                  cs
            Policy Independence
• Choose any one p or choose  (subject to boundary
  conditions)
• No social inefficiency for adding as many distributions as
  you want (true for readers and senders). Each distribution
  can be separately optimized
• Policy can be adjusted to individual senders, not just
  “spammers” and “good-guys”
• Contrast to filter, which suffer from dramatically increased
  type 1 and type 2 errors with more distributions
• There is no incentive to misappropriate the bond ex-post
   Caveats and Adoption Issues
• “bottom chop”
   – Email with high value to the recipient but a low value
     to the sender will not be sent (sender not willing to risk
     bond)
• Infrastructure and Transaction costs
   – requires escrow service(s), server infrastructure
     changes, low cost transaction system for small
     transactions, inter-escrow payment network (later)
• Network effects
   – Mulitiple escrow companies, if not connected, slow
     adoption
      Additional Social Benefits
• Reduced friction; recipients have incentive to publish their
  contact information, not obscure it
• No costly arms race
• Direct Marketing still possible
• Permits communication otherwise eliminated (e.g. political
  speech)
• Costs remain cheaper than post office-style direct
  marketing
• Tailors to an individuals unique preferences
• Signaling information about a recipient via claim history
               Conclusion
• Enabling transactions helps readers more
  than unilateral veto
• Screening mechanism forces senders to
  reveal their type
• Many desirable secondary effects
              Future Work
• Signaling
• Non-uniform distributions
• Allow cs and cr to vary with distribution?
• What happens when sender knows r? (with some
  certainty or exactly)
• Infrastructure issues
• Show determination of minimal  for multiple
  distributions

				
DOCUMENT INFO
Shared By:
Categories:
Tags:
Stats:
views:14
posted:9/3/2011
language:English
pages:51