Slide 1

The Effects of
Switching and Information Costs
on Competition for
Spam Reduction and Email Services
Benjamin Chiao & Jeffrey MacKie-Mason
10/17/05
incur c
Spammers
pay p
Email
Service
Provider
Good & Services
Providers
pass thru
Server
Filters
make type 1 &
2 errors
Email
Clients
Users
Client
Filters
make type 1 &
2 errors
pass thru
Prior Literature
Challenge-Response
(Dwork and Naor, 1993)
incur c
Legal punishment for spammers
(Bazeley, 2003)
Spammers
pay p
Email
Service
Provider
Good & Services
Providers
pass thru
Server
Filters
Content filter, white and
black lists,
(Cranor and LaMacchia,
1998)
make type 1 &
2 errors
Email stamps
(Kraut et al, 2003)
Email
Clients
Users
Client
Filters
Attention-Bond Mechanism
(Loder, Van Alstyne and Wash, 2005)
make type 1 &
2 errors
pass thru
Central Ideas
• To use incentive-based strategies,
embedded in existing markets, to
discourage spam sending, and encourage
better email services
• Our strategies are to decrease the costs of
switching to and the information costs of
finding the best email service providers
(e.g. strongest spam filters)
Already Happening
• Gmail
– Identity-switching
– Double forwarding
Double
Forwarding
Identity
Switching
Identity Switching Relies on
“Spoofing”?
• Senders original identity can still be traced
in the X-headers
– usually not displayed in the simple header
view of most email clients
– any digital signature solutions that prevent
this are not desirable! They might achieve the
opposite effects of not reducing spam by
killing our strategies
Model
• Notations:
– J email service providers
• j~uniform[0,1], filter strength
• θj,, market share
– ns, number of spam sent
– E(nr), expected no. of spam received by users
– c(.), cost function of sending spam, c′>0, and
c′′>0
– p, price for each spam reaching users
• Profit function of a spammer:
• Assumption 1: Spammers care only about
the average filter strength.
• So
• Profit function becomes:
• Zero-profit implies FOC:
Interpretation
• Set p equals to the marginal cost of each
spam times the expected no. of spam
needed for the user to receive one spam.
• Since j~uniform[0,1] implies E()=0.5
What if  is non-stochastic?
Theorem 1
• The number of spam sent and received
will be smaller when the costs of switching
to and finding the smallest  is zero.
Proof
• With no switching and information costs:
– Stochastic :
• where (1) is the smallest order statistics
– Non-stochastic ,
Proof
Other Conjectures (Skipped)
• Multiplier Effect
• Email Services Improvement
To-Do
•
Calibration. To use real data of
- number of email users and accounts
-  and 
to simulate the number of spam reduced
Weaknesses
• The drop of n*s , n*r are continuous but users may care
about discrete thresholds
– But if we also model the fix costs of spammers, n*s , n*r are not
continuous
• Spammers are interested to send different
quantities of spam to different filters. Should we
disclose  then?
Implementation (More Papers?)
• A reputation system to display 
– We may or may not rely on other sources to estimate

– It can even be real time so robots can help users to
switch to the strongest filter real time
• To make the double forwarding and identity
switch common knowledge
– To propose to big email service providers
• To work with standard organizations to minimize
potential conflicts of W3C RFC email headers
etiquette, Yahoo’s Domain Key, PID, Microsoft’s
Sender ID, other authentication system
Extensions
• Spam response rates
• Users care about type 1 and 2 errors
• Users switch only discretely (because of
discrete switch cost) when  decreases.