Filtering Spam with Behavioral Blacklisting
Anirudh Ramachandran, Nick Feamster, and Santosh Vempala
College of Computing, Georgia Tech
801 Atlantic Drive, Atlanta, GA - 30332, USA
{avr,feamster,vempala}@cc.gatech.edu
ABSTRACT
Spam filters often use the reputation of an IP address (or IP ad-
dress range) to classify email senders. This approach worked well
when most spam originated from senders with fixed IP addresses,
but spam today is also sent from IP addresses for which blacklist
maintainers have outdated or inaccurate information (or no infor-
mation at all). Spam campaigns also involve many senders, reduc-
ing the amount of spam any particular IP address sends to a single
domain; this method allows spammers to stay “under the radar”.
The dynamism of any particular IP address begs for blacklisting
techniques that automatically adapt as the senders of spam change.
This paper presents SpamTracker, a spam filtering system that
uses a new technique called behavioral blacklisting to classify email
senders based on their sending behavior rather than their iden-
tity. Spammers cannot evade SpamTracker merely by using “fresh”
IP addresses because blacklisting decisions are based on sending
patterns, which tend to remain more invariant. SpamTracker uses
fast clustering algorithms that react quickly to changes in send-
ing patterns. We evaluate SpamTracker’s ability to classify spam-
mers using email logs for over 115 email domains; we find that
SpamTracker can correctly classify many spammers missed by cur-
rent filtering techniques. Although our current datasets prevent us
from confirming SpamTracker’s ability to completely distinguish
spammers from legitimate senders, our evaluation shows that Spam-
Tracker can identify a significant fraction of spammers that current
IP-based blacklists miss. SpamTracker’s ability to identify spam-
mers before existing blacklists suggests that it can be used in con-
junction with existing techniques (e.g., as an input to greylisting).
SpamTracker is inherently distributed and can be easily replicated;
incorporating it into existing email filtering infrastructures requires
only small modifications to mail server configurations.
Categories and Subject Descriptors: C.2.0 [Computer Communi-
cation Networks]: Security and protection
General Terms: Security, Design, Algorithms
Keywords: spam, botnets, blacklist, security, clustering
1. INTRODUCTION
More than 75% of all email traffic on the Internet is spam [25].
To date, spam-blocking efforts have taken two main approaches:
(1) content-based filtering and (2) IP-based blacklisting. Both of
these techniques are losing their potency as spammers become more
Permission to make digital or hard copies of all or part of this work for
personal or classroom use is granted without fee provided that copies are
not made or distributed for profit or commercial advantage and that copies
bear this notice and the full citation on the first page. To copy otherwise, to
republish, to post on servers or to redistribute to lists, requires prior specific
permission and/or a fee.
CCS’07, October 29–November 2, 2007, Alexandria, Virginia, USA.
Copyright 2007 ACM 978-1-59593-703-2/07/0011 ...$5.00.
agile. To evade content-based filters, spammers have adopted tech-
niques such as image spam and emails explicitly designed to mis-
lead filters that “learn” certain keyword patterns; spammers are also
evading IP-based blacklists with nimble use of the IP address space
(e.g., stealing IP addresses on the same local network [19], stealing
IP address blocks with BGP route hijacking [30]). To make matters
worse, as most spam is now being launched by bots [30], spammers
can send a large volume of spam in aggregate while only sending a
small volume of spam to any single domain from a given IP address.
This “low and slow” spam sending pattern and the ease with
which spammers can quickly change the IP addresses from which
they are sending spam has rendered today’s methods of blacklisting
spamming IP addresses less effective than they once were [11]. For
example, our study in Section 2 shows that, of the spam received
at our spam “traps”, as much as 35% was sent from IP addresses
that were not listed by either Spamhaus [37] or SpamCop [36], two
reputable blacklists. Further, 20% of these IP addresses remained
unlisted even after one month. Most of the IP addresses that were
eventually blacklisted evaded the blacklist for about two weeks, and
some evaded the blacklists for almost two months.
Two characteristics make it difficult for conventional blacklists to
keep pace with spammers’ dynamism. First, existing blacklists are
based on non-persistent identifiers. An IP address does not suffice
as a persistent identifier for a host: many hosts obtain IP addresses
from dynamic address pools, which can cause aliasing both of hosts
(i.e., a single host may assume different IP addresses over time) and
of IP addresses (i.e., a single IP address may represent different
hosts over time). Malicious hosts can steal IP addresses and still
complete TCP connections, which allows spammers to introduce
more dynamism. IP blacklists cannot keep up. Second, information
about email-sending behavior is compartmentalized by domain and
not analyzed across domains. Today, a large fraction of spam comes
from botnets, large groups of compromised machines controlled by
a single entity. With a much larger group of machines at their dis-
posal, spammers now disperse their “jobs” so that each IP address
sends spam at a low rate to any single domain. By doing so, spam-
mers can remain below the radar, since no single domain may deem
any single spamming IP address as suspicious.
IP blacklists must be continually updated to keep pace with cam-
paigns mounted by armies of “fresh” IP addresses. Unfortunately,
a spam campaign may complete by the time the IP addresses are
blacklisted, at which time a new campaign with new IP addresses is
imminent. Blacklisting all new IP addresses is not an option, either:
it creates a nuisance for administrators when legitimate mail relays
are renumbered, as well as for some mobile users.
To keep pace with this dynamism, we propose a new technique
called behavioral blacklisting, which complements existing black-
lists by categorizing spammers based on how they send email, rather
than the IP address (or address range) from which they are sending
it. The intuition behind behavioral blacklisting is that, while IP ad-
dresses are ephemeral as identifiers, spam campaigns, spam lists,

and spamming techniques are more persistent. If we can identify
email-sending patterns that are characteristic of spamming behav-
ior, then we can continue to classify IP addresses as spammers even
as spammers change their IP addresses.
We design a behavioral blacklisting algorithm that uses the set of
target domains that a particular IP address sends mail to as the pri-
mary indicator of its behavior and incorporate this algorithm into a
system called SpamTracker. We use the set of domains that an IP ad-
dress targets within a fixed time window as the feature for clustering
IP addresses that behave similarly. Our clustering algorithm takes as
input an n × d × t tensor, where n is the number of IP addresses
that sent email to any of d domains within one of t time windows.
The algorithm outputs clusters of IP addresses that exhibit similar
sending patterns. Our evaluation of these clusters shows that spam-
ming IP addresses form large clusters that are highly similar to each
other but distinct from the behavior of IP addresses in other clusters.
IP addresses of legitimate senders, on the other hand, do not form
large clusters. SpamTracker can classify a “fresh” IP address as a
spammer or a legitimate sender based on how closely its sending
behavior (i.e., the set of domains that it targets) maps to a clus-
ter that has been marked as known spamming behavior. Using logs
from an organization that manages email for over 115 domains, we
find that SpamTracker detects many spammers before they are listed
in any blacklist, suggesting that SpamTracker can complement to-
day’s IP-based blacklists by catching some spammers earlier than
they would otherwise be caught.
SpamTracker requires little auxiliary information about whether
an email sender is a spammer or a legitimate sender: it takes as in-
put the email-sending patterns of all senders, builds clusters based
on the sending behaviors of (a possibly small set of) known spam-
mers, and classifies each sender based on whether its behavior is
similar to a cluster that resembles known spamming behavior. Un-
like conventional approaches which track individual IP addresses,
SpamTracker tracks behavioral patterns to quickly identify whether
a new IP address exhibits similar patterns to other previously seen
IP addresses. Its ability to track behavior of groups, rather than indi-
vidual IP addresses, allows it to adapt more quickly to ephemeral IP
addresses that may not exhibit strong patterns from the perspective
of any single domain.
Because SpamTracker classifies email based on sending behavior
rather than on more malleable properties of email (e.g., content, or
even IP address), we believe that spammers will have considerably
more difficulty in evading SpamTracker’s classification methods.
Nevertheless, SpamTracker must be agile enough to adapt to spam-
mers’ changing behaviors: spamming patterns (i.e., which domains
are targeted, and how they are targeted) will change over time, and
adversaries that are aware of the SpamTracker algorithm may ad-
just their sending patterns to avoid falling into a particular cluster.
We believe, however, that automated, large-scale behavior such as
spamming will always give rise to clustering, and the challenge is to
design SpamTracker to adapt the clusters it uses for classification,
even as the spammers themselves attempt to evade them.
The paper is organized as follows. Section 2 motivates behav-
ioral blacklisting. Section 3 presents a brief background on cluster-
ing techniques and describes EigenCluster [7], the clustering algo-
rithm that we use in SpamTracker. Section 4 describes the design
and implementation of SpamTracker, and Section 5 presents our
validation results and compares the performance of SpamTracker
to state-of-the-art IP-based blacklists and spam trap deployments.
In Section 6, we discuss various extensions of SpamTracker and
deployment-related concerns. Section 7 presents related work, and
Section 8 concludes.
2. MOTIVATION
This section provides background on current email spamming
practices and the performance of blacklists. In Section 2.1, we
1e-04
0.001
0.01
0.1
1
0 5 10 15 20 25 30
Fr
ac
tio
n
of
U
ns
ee
n
IP
s
Time (days)
Not seen for two months
Not seen for 1 month
Not seen for 1 day
Figure 1: Fraction of spamming IP addresses that were not ob-
served at any of 115 domains for the past 1 day, past month, and
past 2 months.
present the volumes and rates at which IP addresses in our traces
send spam to each domain; we find that spammers exhibit send-
ing patterns that make it difficult to reliably detect and track spam-
ming IP addresses. In Section 2.2, we provide background on cur-
rent IP-based blacklisting techniques (e.g., DNS-based blacklists)
and present a study of their effectiveness.
2.1 The Behavior of Spamming IP Addresses
We present statistics on the network-level behavior of spammers,
focusing on the techniques that make building the reputation of any
particular IP address difficult. We study two aspects in particular:
(1) Persistence: How much spam does a particular IP address send
in a day, and how does the set of IP addresses change over time?
(2) Distribution: What is the distribution of spam across target do-
mains for any particular IP address, and how does this distribution
change over time?
2.1.1 Persistence: “New” IP addresses every day
To determine the extent to which spamming IP addresses remain
stable, we study the IP addresses that send spam to over 115 distinct
domains, which collectively received 33 million pieces of spam dur-
ing March 2007.1
Figure 1 shows the number of “new” IP addresses that these do-
mains observed per day over the course of a month. The top line
shows the fraction of IP addresses that were seen in the trace for
a particular day that were never seen before in the trace (other
lines show fraction of spam from IP addresses that appeared on
the immediately preceding day, or within the month). Indeed, spam
is coming from different IP addresses every day, and about 10%
of IP addresses seen on any particular day were never seen be-
fore at any of the target domains. Thus, even given perfect mecha-
nisms for maintaining reputation about email senders and relatively
widespread observation, a significant number of IP addresses that
have never been seen before are sending spam on any given day.
Lack of persistence in spamming IP addresses makes maintaining
reputation about spammers based solely on IP addresses difficult,
since the blacklisted IP addresses keep changing. Given no previ-
ous information about the activity of an IP address, a conventional
blacklist will not be able to reliably block spam from that address.
2.1.2 Distribution: Some IPs target many domains
Existing blacklisting techniques collect reputation information
about spam or spam senders based on the activity observed at a sin-
gle domain (e.g., if a spammer sends a significant amount of spam
to a single IP address, if it hits a spam trap, etc.) [36, 37]. Although
1Section 5.1 describes this dataset (as well as the others that we
used in our evaluation) in more detail.

cant fraction of spam from senders that successfully evade conven-
tional blacklisting techniques.
Experiment 2: Are accepted senders blacklisted later? The sec-
ond set of logs are from an organization that hosts email service
for over 700 domains, about 85 of which were active during March
2007 (our observation period). This provider’s mail servers reject
or accept email based on a combination of techniques, including
multiple blacklist lookups (Spamhaus [37], SORBS [34], etc.) and
a large collection of customized heuristics. This provider blocks up
to twice as much spam as any single blacklist.
Using our daily snapshot of the Spamhaus blacklist as a basis
for comparison, we study the effectiveness of this email provider’s
blocking heuristics by determining the fraction of mail that the
provider accepts. Our results show that even this provider’s ad-
vanced filtering does not ensnare a significant collection of spam-
mers: Of the 5,084,771 senders that passed the spam filters, only
110,542 (2%) became listed in the Spamhaus blacklist during the
following month. This fraction is significantly lower than the 15%
quoted by this provider as the fraction of accepted email that is later
reported as spam, which suggests that current blacklists remain in-
complete, even after long periods of time.
2.2.3 Responsiveness
Many DNSBLs do not list an IP address before they receive mul-
tiple end-user reports about a spam sender; some even perform
manual verification. Meticulous verification can reduce the likeli-
hood of blacklisting “good” senders, but doing so also limits re-
sponsiveness. In this section, we quantify the responsiveness of the
Spamhaus DNSBL by determining, for the IP addresses that were
eventually listed in April 2007, how long those IP addresses were
active before they eventually were blacklisted.
As before, we use snapshots of the Spamhaus blacklist, but we
also use hourly “diffs” of the blacklist to determine when a new IP
address was added. We examine email logs from March 1–31, 2007
and blacklist data from April 1–30, 2007. For each IP address that
was not listed when it first sent spam to one of our spam traps but
was eventually listed at some later point in April 2007, we compute
the delay between the first occurrence of the IP at our trace to the
first time that the IP address became listed in Spamhaus.3
Even when blacklists do list spamming IP addresses, the process
of updating the blacklist may be slow. Figure 3 shows the time-to-
listing for all IPs that were unlisted during the receipt of the email
but eventually appeared at the blacklist in April 2007. In the case
of the spam traps, 10–15% of spam senders that were unlisted at
receipt of spam remained so 30 days after spam was received. The
fraction is a strong indicator of the sluggishness of blacklists, be-
cause sending email to a spam trap automatically labels the sender
as a spammer. In the case of the provider that serves millions of
real customers (“Organization”), almost 20% of senders that were
unlisted when email was received remain unlisted for over 30 days
before eventually appearing in the blacklist.
This analysis indicates that reactive blacklists are sometimes
slow to respond, even for confirmed spammers; this slow re-
sponsiveness, coupled with the ability to continually send spam
from “fresh” IP addresses (Section 2.1.1) represents a significant
“window of opportunity” for spammers to send spam from non-
blacklisted IPs. Motivated by this slow responsiveness, the next
section proposes a complementary approach to blacklisting that is
3Because we only have the Spamhaus database for April, we can-
not determine the exact listing time for IP addresses that were in
the database on April 1, 2007; rather, we only know that they were
listed between the time the spam was observed in March and April
1, 2007 (“less than 30 days” in Figure 3). If the IP address was not
listed by April 1, 2007, we assume that whenever the IP becomes
listed in April is the first time Spamhaus listed it. This assumption
is reasonable as Spamhaus lists persistent spammers for a minimum
of 30 days [1].
0.8
0.82
0.84
0.86
0.88
0.9
0.92
0.94
0.96
0.98
1
60504030Less than days
Cu
m
ul
at
iv
e
Fr
ac
tio
n
of
E
ve
nt
ua
lly
B
la
ck
lis
te
d
IP
s
Days to Listing
Spamtrap 1
Spamtrap 2
Organization
Figure 3: Time-to-listing for the Spamhaus blacklist for IP ad-
dresses that were unlisted at the time spam was received, but
were eventually blacklisted. Note: y-axis starts at 0.8.
based on email sending patterns, rather than the reputation of an IP
address alone.
2.3 The Case for Behavioral Blacklisting
Although individual IP addresses’ sending behavior may change
across time, we posit that (1) the sending patterns exhibited
by spammers are sufficiently different from those of legitimate
senders; and (2) those patterns become more evident when email
senders can be observed across many receiving domains. Based
on these two hypotheses, the rest of the paper proposes a sys-
tem called SpamTracker, which proactively blacklists email senders
based on the set of domains they target. SpamTracker relies on a
technique that we call behavioral blacklisting, which attempts to
classify based on their network behavior, rather than their identity
or the contents of the emails they send. While individual IP ad-
dresses may be ephemeral, they may exhibit “familiar” spamming
patterns (i.e., similar to those of already well-known spamming IP
addresses) that become evident when sending patterns are observed
across multiple domains.
3. CLUSTERING ALGORITHM
SpamTracker uses a spectral clustering algorithm proposed and
analyzed by Kannan et al. [18] and made efficient in practice by
Cheng et al. [7]. Section 3.1 presents an overview of the spectral
clustering approach, and Section 3.2 describes how we apply spec-
tral clustering within SpamTracker.
3.1 Spectral Clustering
Spectral clustering refers to partitioning algorithms that rely on
the principal components of the input. There are generally two basic
variants which can be viewed as (a) one-shot or (b) recursive. Given
an object-feature matrix A with the goal of clustering the objects
(rows) of A, a one-shot algorithm would find the top few singular
vectors of A (say k) and either project to their span or create a
cluster for each one by assigning each row to that vector in which
it has the largest component. A recursive algorithm, on the other
hand, uses one singular vector to partition the rows and recurses on
the two parts. We focus on this type of algorithm.
The method in Cheng et al. [7] (“EigenCluster”) has two phases:
a top-down divide phase and a bottom-up merge phase. In the divide
phase, the algorithm normalizes a given nonnegative input matrix so
that all rows have the same sum, then computes the second largest
right singular vector. It sorts the rows according to their compo-
nents in this vector and partitions this sequence at the point where
the corresponding cut has minimum conductance (among the n− 1
possible cuts of the sequence). The conductance of a partition is the
total weight of entries across the partition divided by the smaller

Figure 4: An IP× IP matrix of related spam senders; IP ad-
dresses that send mail to similar sets of domains are grouped
into distinct clusters; the intensity of a pixel at (i,j) indicates i’s
similarity to j.
of the total weights incident to each side [18, 33]. After finding
the partition, it recurses on each side until only singletons remain.
This completes the divide phase, whose end result is a tree (the root
represents all the rows, the leaves are individual rows). The merge
phase finds a tree-respecting partition, i.e., one where every clus-
ter corresponds to the entire subtree attached at some node of the
tree. For many objective functions, it does this by dynamic pro-
gramming, in a bottom-up fashion. The specific function we use for
the merge phase is called correlation clustering [7].
3.2 SpamTracker: Clustering Email Senders
SpamTracker classifies an email sender purely based on its send-
ing behavior, ignoring content and variable handles for classifica-
tion such as dynamically-allocated IP addresses. The intuition be-
hind SpamTracker is that sending patterns of spamming hosts are
similar to other senders and remain relatively stable, even as the
IP addresses (or actual systems) that are sending the emails change.
Consider the case of a spamming bot: Whatever the particular spam-
ming behavior of a spamming bot, it is likely to be similar to other
bots in its own botnet. Because botmasters in large botnets have
only coarse-grained control over their bots [26], spamming patterns
of bots will typically be similar across targeted domains even if each
bot sends low volumes of spam to each domain. Thus, clustering
spammers based on their sending patterns provides a way for their
early detection, irrespective of their particular identities (e.g., the IP
address) or blacklisting status. It follows from the above that, spam
sent from even a newly-enlisted bot (i.e., from an IP address that
has not been observed to send spam) will likely be caught by Spam-
Tracker because its behavior will cluster it with other known bots
in the botnet.
The SpamTracker algorithm proceeds in two stages: (1) clus-
tering and (2) classification. In the unsupervised clustering stage,
SpamTracker accepts as input a n × d × t tensor M , where n is
the number of IP addresses that sent email to any of d domains
within any of t particular time windows. Thus, M(i, j, k) denotes
the number of times IP address i sent email to domain j in time
slot k. SpamTracker first collapses the time axis to obtain an n × d
matrix M ′:
M ′(i, j) =
t
X
k=1
M(i, j, k).
It clusters the matrix M ′ using the spectral clustering algorithm
described in Section 3.1. The output of the clustering stage is the set
of clusters of IP addresses C = C1, C2, . . . , Ck, where ∪ki=1Ci =
IPs in M and Ci ∩ Cj = φ for i 6= j. Logically, the set C consists
of groups of IPs in M that have similar behavior in their target
domains. Each cluster is associated with a traffic pattern, obtained
by averaging the rows corresponding to IPs that fall in the cluster.
For a cluster c, we call this vector cavg .
cavg =
P|c|
i=1 M
′
c(i)
|c|
where M ′c(i) is the submatrix comprising the rows of cluster c. In
the classification stage, SpamTracker accepts a 1 × d vector r that
corresponds to the recent behavior of an IP. It then calculates a score
S(r) for this queried IP address using the following equation.
sim(r, c) = r · cavg|cavg |
(1)
Intuitively, sim(r, c) measures the similarity of the row vector r to
cluster c by performing an inner product of r with the normalized
average of rows in cluster c. A cluster that has a similar set of target
domains as r would have a large inner product.
We calculate the spam score S(r) as the maximum similarity of
r with any of the clusters.
S(r) = max
c
sim(r, c). (2)
S can be used to filter or greylist (i.e., temporarily reject with the
assumption that a legitimate mail sender will eventually retry) spam
by a mail service provider at or before the SMTP dialogue stage. We
set a threshold such that if the row for an IP that is looked up has
score higher than the threshold, it is flagged as spam. The threshold
can be different for each cluster.
Querying an IP address is inexpensive: only Equations 1 and 2
need to be computed per lookup. The next section explains the de-
sign of SpamTracker in detail and the optimizations we use to im-
prove the lookup speed and the overall robustness of the system.
4. DESIGN
This section describes how SpamTracker can be integrated into
an existing email infrastructure. We present a brief overview of the
system and then describe in detail its two basic operations: (1) com-
puting the clusters that form the basis of the classifier; and (2) clas-
sifying a new IP address when it arrives.
4.1 Overview
The spectral clustering algorithm in Section 3.2 serves as the
back-end of SpamTracker. The behavioral classifier that accepts
lookups from mail servers and assigns scores to the queried senders
forms the front-end. Figure 5 shows the high-level design of Spam-
Tracker and the interaction between the back-end (which performs
clustering and classification operations) and the interface to mail
servers (which receives email sending patterns as input to the clus-
tering algorithm and answers queries about the status of any partic-
ular IP address); to an ordinary mail server, the interface to Spam-
Tracker looks like any other DNS-based blacklist, which has the
advantage that existing mail servers need only to be reconfigured to
incorporate SpamTracker into spam filtering decisions. We discuss
how SpamTracker can be incorporated into existing infrastructure
in Section 6.2.
SpamTracker’s clustering algorithms rely on the assumption that
the set of domains that each spammer targets is often more stable
than the IP addresses of machines that the spammer uses to send the
mail. Rather than maintaining reputations of senders according to
their IP addresses, SpamTracker uses the vector representing how
a sender sends traffic across domains, r, as a “behavioral finger-
print” and determines whether this fingerprint resembles a known
spamming cluster. Section 4.2 describes how SpamTracker builds
clusters of known spammers, and Section 4.3 explains how Spam-
Tracker determines whether an email sender’s sending patterns re-
semble one of these clusters.

4.2 Clustering
SpamTracker uses the spectral clustering algorithm from Sec-
tion 3.1 to construct the initial set of clusters. SpamTracker’s clus-
tering takes as input email sending patterns about confirmed spam-
mers (i.e., the volume of email that each confirmed spamming IP
address sends across some set of domains) over some time window
to construct the matrix M(i, j, k). This input requires two compo-
nents: (1) an initial “seed list” of bad IP addresses; and (2) email
sending patterns for those IP addresses. This section describes in
turn how SpamTracker might be able to acquire this type of data.
Data about spamming IP addresses is easy to obtain, and Spam-
Tracker could use any such initial list of IP addresses to “bootstrap”
its initial clusters. For example, an Internet Service Provider (ISP)
that uses conventional SpamAssassin [35] filters to filter spam could
use that list of IP addresses as its initial spammer IP addresses to be
used for the basis for clustering.
The sending patterns of each of the spamming IP addresses is
more difficult to obtain because it requires visibility into the emails
that many domains have received. Our evaluation of SpamTracker
(Section 5) uses an email hosting provider’s decisions about early
mail rejects from hundreds of domains to compute these clusters,
but, in practice, other systems like SpamTracker could also likely
gain access to such data.
To build the rows in M for each spamming IP address, participat-
ing domains could submit IP addresses that they have confirmed to
be spammers as they do with blacklists, but based on our findings
of the “low and slow” sending patterns of spammers (Section 2),
SpamTracker will be most effective if it maintains sending patterns
across domains for as many IP addresses as possible and subse-
quently clusters based on some subset of those that are labelled
as spam by at least one domain. Fortunately, SpamTracker could
obtain these sending patterns from receiving mail servers’ queries
to the classifier4, at least from some subset of trusted domains.5
Specifically, a lookup for IP address a from domain d is a reason-
able indicator that a has sent email to d, so SpamTracker can build
vectors for all such addresses a and later build the matrix M from
just those addresses that are confirmed to be spammers.
4.3 Classification
SpamTracker maintains a vector representing the sending pattern,
r, for each IP address a compiled from reports from the mail servers
of participating domains. SpamTracker collects these sending pat-
terns as mail servers from trusted participating domains perform
lookups to SpamTracker on address a, using the same method for
collecting these patterns for all IP addresses during the clustering
phase (described in Section 4.2).
Given an r for some IP address a, SpamTracker returns a score
S(r) (computed using Equation 2, Section 3.2) whose magni-
tude determines how closely this fingerprint resembles a confirmed
spamming pattern (i.e., cluster). SpamTracker can simply return
S(r) to the querying mail server, which can then incorporate this
score into its existing mail filtering rules. An important benefit of
the classification process is that S(r) can be computed using only
an IP address’s r vector and the cavg rows for the spam clusters,
both of which can be replicated and distributed (providing robust-
ness against attack, as well as load balance). Clustering requires
‘r’ vectors from as many IP addresses as possible; even though
it requires aggregating sending information from many sending
domains (and, hence, from potentially many SpamTracker repli-
4Note that the query mechanism needs a way of finding the email
domain name of the organization performing the query. DNS re-
verse lookups, or extra information in the query packets, could pro-
vide such a mechanism.
5Because previous work has observed that bots occasionally per-
form reconnaissance queries against blacklists [29], we cannot as-
sume that all queries to the blacklist reflect the receipt of email by
a mail server.
Figure 5: The high-level design of SpamTracker. The cluster-
ing component of SpamTracker accepts information about email
senders as an IP×domain×time tensor and computes clusters
of related senders (and corresponding average vectors). The
classification component accepts queries for IP addresses and
returns a score, S(r), for the IP’s behavior.
cas), this aggregation and clustering can be performed on a slower
timescale than classification.
4.4 Tracking Changes in Sending Patterns
SpamTracker must recompute new clusters as sending patterns
change. Our implementation of SpamTracker reclusters at fixed
intervals, but in practice SpamTracker might only recluster when
sending patterns no longer map to any existing clusters. Re-
clustering cost (time, memory, CPU) increases with larger input ma-
trices, so clustering on very large time windows may be impractical.
We use an efficient re-clustering method that preserves historical in-
formation but keeps clustering cost approximately constant. At the
beginning of each clustering phase, we add all average rows from
the previous clustering stage scaled by the size of the cluster each
row represents, which produces the effect of clustering on the input
of both stages without the added cost.
5. EVALUATION
This section describes the evaluation of SpamTracker. In a real
deployment, SpamTracker could compute clusters based on send-
ing patterns across many domains for some time interval. To emu-
late this scenario, we construct the SpamTracker classifier by con-
structing M(i, j, k) from the email logs of a large organization that
manages mail servers for hundreds of domains. We use the matrix
for time window at [t, t + ∆t) to build the classifier, and the data
in the window [t + ∆t, t + 2 ∆t) to validate our classification.
Section 5.1 summarizes the data sets used in our evaluation. Sec-
tion 5.2 describes the properties of the resulting clusters and the
validation results, and Section 5.3 describes our evaluation of Spam-
Tracker’s ability to improve upon existing blacklisting and blocking
techniques by classifying spammers ahead of blacklists.
5.1 Data
Table 2 summarizes the traces, their duration, and the data fields
each trace provides. Our primary data is a set of email logs from

Trace Date Range Fields
Organization Mar. 1 – 31, 2007 Received time, remote IP,
targeted domain, whether
rejected
Blacklist Apr. 1 – 30, 2007 IP address (or range), time
of listing
Table 2: Data sets used in evaluation.
a provider (“Organization”) that hosts and manages mail servers
for over 115 domains. The trace also contains an indication of
whether it rejected the SMTP connection or not. We also use the full
database of Spamhaus [37] for one month, including all additions
that happened within the month (“Blacklist”), to help us evaluate
the performance of SpamTracker relative to existing blacklists. We
choose the Blacklist traces for the time period immediately after the
email traces end so that we can discover the first time an IP address,
unlisted at the time email from it observed in the Organization trace,
was added to Blacklist trace.
5.2 Clustering and Classification
To study the properties of the clusters that SpamTracker com-
putes, we build the SpamTracker classifier using data for a window
∆t at time t, and use it to assign a spam score S(r) senders in the
window [t + ∆t, t + 2 ∆t). We set ∆t to be 6 hours; clustering
using different time intervals (which we intend to explore in future
work) may also help SpamTracker perform better.
Figure 6(a) shows the distribution of these scores for all IP ad-
dresses in a 6-hour window, separated into two plots based on
whether the Organization decided to reject the mail early or accept
it for delivery. A high score implies that the sending pattern for the
classified IP is similar to a known spamming pattern. The low-score
region (where S(r) < 1) comprises IP addresses whose patterns are
unknown to the classifier. Senders that map into this range should
not necessarily be considered legitimate; rather they simply do not
have a recognized, blacklisted sending pattern. High scores reflect
IP addresses whose sending patterns are very similar to the aver-
age rows of the classifier. As expected, the distribution of mails
rejected by the organization tend towards larger values of S(r). We
suspect that because legitimate email senders likely will not mimic
each other’s sending patterns, the IP addresses in this region—both
in the “accepted” and “rejected” plots—are likely to contain spam-
mers. Indeed, in Section 5.3, we show that SpamTracker correctly
classified IP addresses in that were accepted by the Organization
but were eventually blacklisted.
Ideally, users of SpamTracker should be able to set a single
threshold for S(r) that clearly separates the majority of legitimate
email from the majority of spam, but setting a single threshold for
the experiment shown in Figure 6(a) could result in misclassify-
ing a large fraction of received mail. For example, though setting a
threshold of 10 would blacklist only about 5% of the Organization’s
accepted mail, it would only correctly classify 10% of all of the re-
jected mail. In fact, a lower threshold may be more appropriate: as
we describe in Section 5.3 below, a significant fraction of accepted
mail is still spam, and, in many cases, SpamTracker captures this
spam before the Organization or Spamhaus does. However, without
ground truth data, it is difficult to determine a precise false positive
rate, because “accepted” mail may simply be misclassified spam.
We believe that the quality of data (rather than the classification
algorithm itself) is affecting our ability to separate the accepted
and rejected mail with a single spam score. First, the data set is
not cleanly labelled: the decisions of the Organization concerning
whether to accept or reject a mail are not in fact a ground truth in-
dicator as to whether mail is legitimate: The Organization estimates
that as much as 15% of accepted mail is spam, and, as we show in
Section 5.3, the emails that were accepted by the Organization for
which SpamTracker assigns high scores may in fact be undetected
spammers. Second, SpamTracker performs best when the represen-
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.001 0.01 0.1 1 10 100 1000
CD
F
Score
Sending to a single domain
Unseen multi-domain
sending patterns
Seen multi-domain
sending patterns
Accepted
Rejected
(a) For all IP addresses.
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
0.001 0.01 0.1 1 10 100 1000
CD
F
Score
Accepted
Rejected
(b) For IP addresses with maximum similarity to some
cluster with a distributed sending pattern.
Figure 6: Score distribution for SpamTracker’s classification for
(a) All IPs in a “round” of classification, and (b) IPs that have
maximum similarity with a cluster whos cavg is not dominated
by a single column. Evaluation for a 6-hour period using a clas-
sifier trained using the previous 6-hour window.
tative sending behavior for a cluster is distributed across multiple
domains, rather than concentrated in a single domain. Figure 6(a)
shows that many emails have a spam score of 1, which implies that
the classified IP address’s pattern is similar to a cluster whose av-
erage row is dominant in one column. According to Equations 1
and 2, this pattern will return a similarity of about |r|. Because, in
our dataset, a majority of senders in most small time windows send
email to only a single domain, |r| is 1 for 50% of accepted email
and 30% of rejected email. Our dataset often has email senders that
send mail to only a single domain in a time window.
Figure 6.5 shows the distribution of S(r) for IP addresses that
have maximum similarity with a single cluster whose cavg is not
dominated by a single column. The “accepted” and “rejected” dis-
tributions separate more cleanly because legitimate IP addresses
that have maximum similarity with this cluster will likely not have
sent mail to all domains comprising the average row of this cluster
(although the spammers in this cluster will likely hit all or most do-
mains). A better distribution of monitors might result in a more even
observation of sending patterns, which should result in a distribu-
tion of S(r) that more closely resembles that shown in Figure 6.5.
5.3 Detecting “New” Spammers
To estimate the fraction of spammers that SpamTracker’s clus-
tering techniques can detect in advance of conventional blacklisting
techniques, we study a subset of the email with the highest spam
scores and determine whether any emails from this subset were
eventually reported as spam (e.g., by users, or some other aux-
iliary technique). Operators at the Organization acknowledge that

about 15% of email that is initially accepted falls into this category.
To estimate how well SpamTracker would perform at catching this
15% of misclassified mail, we examine the 620 emails were initially
missed by the Organization’s filters but eventually blacklisted in the
following month. Of these, 65 emails (about 10%) had a spam score
of S(r) > 5 (from Figure 6(a)), suggesting that SpamTracker could
complement existing filtering mechanisms by capturing additional
spam that existing filters miss.
6. DISCUSSION
Although the general method of behavioral blacklisting shows
promise for fast classification of spammers, there is much room
for improvement, particularly with respect to the classification al-
gorithms (which could, for example, incorporate other features as
input). This section proposes specific areas where the classifica-
tion algorithms could be improved, surveys how filtering techniques
based on behavioral blacklisting could ultimately be deployed in an
operational network, and presents our ongoing efforts to do so. We
also discuss how behavioral blacklisting scores might be integrated
into existing spam filtering systems and some of the issues that may
arise in implementation and deployment.
6.1 Improving Classification
IP addresses that are most similar to a single spamming cluster
can be classified more accurately. In order to achieve this separation
for all new IPs, we propose two improvements to SpamTracker that
may result in better clusters.
Using more features for clustering. Although SpamTracker uses
target domains to construct the initial object-feature matrix (Sec-
tion 3.2), other behavioral features may be able to better classify
spammers. Temporal patterns such as the time interval between suc-
cessive emails received from an IP (or alternatively, the sending fre-
quency of the IP) is one such feature. Botmasters often manage all
their bots using unified interfaces that may also be used to dissemi-
nate spam templates and mailing lists to bots [26], so these bots may
exhibit similar temporal behavior (perhaps spamming frequencies)
in addition to their similarity in target domains.
Improved similarity computation. In Equation 1, all columns of
IP’s “fingerprint” vector, r, are weighted equally. Some domains
may be better at distinguishing one cluster of spammers from an-
other. For example, spammers targeting victims in different coun-
tries may send email to country specific domains as well as to ubiq-
uitous domains (e.g., gmail.com). In this case, the country-specific
domains may be more helpful in distinguishing the two sets of
spammers. Our ongoing work includes experimenting with an al-
gorithm that weights each column (domain) differently.
6.2 Incorporating with Existing Systems
We discuss how SpamTracker can be incorporated to comple-
ment the existing deployments of mail servers and spam filters.
We describe two possibilities below: integration with existing filters
and on the wire deployment. In either case, the back-end of Spam-
Tracker can remain the same: it only needs to run a DNS server (or
another popular query interface such as XML-RPC) that accepts re-
quests for IP addresses, retrieves the classification score S(r) from
the SpamTracker classification engine, and returns the score to the
client. In this sense, SpamTracker is a stand-alone system that can
even be used internally within an organization.
Option 1: Integration with existing infrastructure. SpamTracker
could be incorporated into existing filtering systems on mail servers
by providing an additional “confidence score” for these filters that
help them determine whether a particular piece of email is spam in
terms of sender behavior. Because SpamTracker provides a simple
interface (i.e., it takes as input an IP address and returns a score),
it can be incorporated into any existing spam filtering engine (e.g.,
SpamAssassin [35], MailAvenger [23]) in the same way that any
other blacklist information would be added as a filtering criterion.
Using this system would be easy: the addition of one line to the
configuration of most mail filtering software should allow users to
benefit from SpamTracker’s filtering strategy.
The disadvantage, however, is that it does not stop email traffic
close to the source: the mail server that receives the spam drops
the mail only after the traffic has already traversed the network and
consumed resources on the receiving mail server.
Option 2: “On the wire” deployment. Unlike most existing spam
filtering or classification systems, SpamTracker has the unique ad-
vantage that it can classify email senders solely based on the source
IP address and destination domain of the mail being sent (i.e., it
does not require examining or analyzing an email’s contents). Thus,
another possibility for deploying SpamTracker involves deploying a
network element that can examine traffic “on the wire” and identify
connections to mail servers from IP addresses that fall into clusters
with high spam scores. Such a system could be deployed anywhere
in the network, not just at the receiving mail server.
The disadvantage to this strategy is that deployment involves sev-
eral additional steps: in particular, such a filtering element would
need a channel to receive up-to-date information about both the
email sending clusters (i.e., their average vectors, and their “spam-
miness”) and the vector for any particular sending IP address (i.e.,
to which domains it has sent). Maintaining up-to-date information
about clusters and sending IP addresses in such a distributed, dy-
namic setting may prove challenging in practice.
6.3 Deployment Challenges
A SpamTracker deployment must be scalable (i.e., it must be able
to handle a large volume of email and a large number of senders)
and robust (i.e., it must be resistant to attack and remain highly
available). To achieve these goals, we believe that SpamTracker
could ultimately be distributed: many servers (possibly the same
ones who manage mail for various domains) report sender behavior
to a centralized location that performs the clustering. SpamTracker
must aggregate data from many domains, compute the correspond-
ing clusters of email senders, and return scores from many sources;
in doing so, it faces scalability and reliability challenges that could
be addressed with the following enhancements.
Better scalability with data compression. SpamTracker’s cluster-
ing algorithm is centralized, which raises scalability concerns, both
for bandwidth (to exchange information between domains) and in
terms of processing power (clustering complexity increases with in-
put size). We are investigating ways to reduce load by distributing
the clustering process. For example, compressing cluster informa-
tion into average rows before sending this information to a cen-
tralized server may reduce bandwidth consumption: SpamTracker
requires the full IP × domain matrix from each source to perform
clustering, but requires only the average row vectors for each cluster
(i.e., the output of the algorithm) for classification.
Better reliability with replication and anycast. To improve avail-
ability, SpamTracker servers could be replicated and placed in dif-
ferent locations or on independent networks. Multiple servers might
be anycasted or managed by different organizations (much like the
DNS root nameserver infrastructure today), all of which perform
the same computation and disseminate average rows to second-level
servers, which in turn respond to user lookups.
6.4 Evasion
SpamTracker must be resistant to attacks that mislead the clus-
tering engine in ways that can cause spam to be misclassified as
legitimate email, and vice versa. To improve classification robust-
ness, SpamTracker could form clusters based on email sending pat-
terns from a smaller number of trusted email recipients (e.g., a

few hundred trusted domains), each of which communicates with
the SpamTracker system over a secure channel. Although Spam-
Tracker’s clustering benefits from more inputs about email senders,
it can serve as a classifier for a much larger set of domains that it
does not necessarily trust to provide data for forming the clusters.
If spamming bots in a botnet select target domains from the same
distribution, SpamTracker’s clustering algorithm will include these
spammers in the same cluster. Still, SpamTracker is limited by the
time window used for clustering (e.g., 6 hours, as in Section 5), and
a spammer might exploit this weakness to evade SpamTracker. We
are improving SpamTracker to automatically adjust the window in
response to the fraction of received email in the last window that
was classified as spam. The intuition is that the fraction of spam
does not change much over short timeframes, and a decrease in the
fraction of flagged email indicates that the window is too small to
cluster similar IPs together. Spamming bots might also try to emu-
late the distribution of target domains (or other behavioral features)
of normal senders, but by doing so, they may be inherently less
effective (e.g., they may have to reduce their sending rate or the
expansiveness of their target list).
6.5 Sensor Placement
A set of domains that observes more even sending behavior
across domains may be able to better distinguish spammers from
legitimate senders. Recall from Section 2.1.2 that 90% of the spam
we observe is received by only 84 of the 115 domains from which
we observe email, and that only about 15% of the senders in our
traces target more than one of the domains from which we can ob-
serve sending patterns at the email hosting provider. Based on our
experiments using only clusters where the average vectors are less
“skewed” towards a single domain (Figure ), we expect that a more
even distribution of sensors email would further improve the Spam-
Tracker classifier. Many commercial spam filtering companies (e.g.,
IronPort [16], Secure Computing [31]) may already have this data.
Another option for sensors would be ubiquitous Web mail domains
such as hotmail.com, gmail.com, etc.
7. RELATED WORK
In this section, we discuss several areas of related work: (1) Pre-
vious characterization studies, several of which offer statistics that
help build the case for behavioral blacklisting; (2) Existing systems
for spam filtering, many of which use distributed monitoring but
incorporate different algorithms for classification; (3) Previous ap-
proaches for classifying email into legitimate email and spam.
Blacklisting and identity. SpamTracker relates to previous black-
listing proposals. Conventional blacklists constitute lists of IP ad-
dresses of likely spammers and are intended to help spam filters [15,
23, 35] make better decisions about whether to block a piece of
email based on the sender. Some blacklists are policy-based (e.g.,
they list all IP addresses that belong to a certain class, such as di-
alup addresses [34]). Other IP-based blacklists are “reactive”: they
attempt to keep track of whether an IP address is a spammer, bot,
phisher, etc. and keep this list up-to-date as hosts are renumbered,
botnets move, and so forth [24, 36, 37, 39]. These blacklists essen-
tially maintain lists of IP addresses and must be vigilantly main-
tained so as to not going out of date. Sender Policy Framework
(SPF) attempts to prevent IP addresses from sending mail on behalf
of a domain for which they are not authorized to send mail [42], and
domain keys associate a responsible identity with each mail [3]. Al-
though both frameworks make it more difficult for an arbitrary IP
address to send mail, they do not allow a recipient to classify an
email sender with an unknown reputation.
Collaborative filtering and whitelisting. SpamTracker resembles
the many existing systems that take inputs from many distributed
sources to build information about known spam (or spammers).
Some of the most widely deployed collaborative filtering systems
characterize known spam based on the contents of a piece of spam
that was reported or submitted by another user or mail server [10,
12, 20, 27, 28, 40]. These systems allow mail servers to compare the
contents of an arriving piece of email to the contents of an email that
has been confirmed as spam; they do not incorporate any informa-
tion about network-level behavior.
Other systems collect information from distributed sets of users
either to help filter spam or decrease the probability that legiti-
mate mail is mistakenly filtered. IronPort [17] and Secure Comput-
ing [32] sell spam filtering appliances to domains which then pass
information about both legitimate mail and spam back to a central
processing engine that in turn improves the filters. The widespread
deployment of these products and systems make them ideal candi-
dates for the deployment of an algorithm like SpamTracker.
Characterization studies. Our recent characterization study of the
network-level behavior of spammers observes spamming behavior
from the perspective of a single spam “trap” domain [30]. In this
study, we observed that any particular IP address sends only a small
volume of spam to the particular domain being observed over the
course of 18 months. Duan et al. recently performed a similar study
that observes similar characteristics [13]. Our characterization of
spammers in Section 2 builds on these previous studies by observ-
ing email sending patterns across domains and time.
Content-independent blocking. Like SpamTracker, Clayton’s
“spamHINTS” project also aims to characterize and classify spam
with analysis of network traffic patterns, rather than email con-
tents [38]. Earlier work on “extrusion detection” involves instru-
menting a mail server with a log processing program to detect
senders of spam both at the local ISP [8] and in remote ISPs [9]. Al-
though Clayton’s proposed methods are similar in spirit to our work
(in that the methods rely on examining traffic patterns to distinguish
legitimate email senders from spammers), the methods generally
involve heuristics related to SMTP sessions from a single sender
(e.g., variations in HELO messages, attempt to contact incoming
mail servers to send outgoing mail); in contrast, SpamTracker relies
on a wider deployment of traffic monitors (i.e., it relies on observ-
ing email sending patterns from many domains) but is then able to
for more protocol agnostic “fingerprints” for email senders that are
likely spammers. Trinity counts email volumes to identify emails
that are likely sent from bots [6]; it could also be used to track email
sending patterns for input to SpamTracker.
Clustering for spam classification. Previous studies have at-
tempted to cluster spammers based on an emails contents, such as
the URLs contained in the bodies of the emails [4, 22]. Li et al.
focus on clustering spam senders to predict whether a known spam-
mer will send spam in the future [22], and Anderson et al. cluster
spam according to URLs to better understand the relationship be-
tween the senders spam messages that advertise phishing and scam
sites and the Web servers that host the scams themselves [4]. These
systems cluster emails based on content, while SpamTracker clus-
ters email senders based on their sending behavior. Unlike the meth-
ods of Li et al., SpamTracker’s clustering techniques can also clas-
sifying previously unseen IP addresses.
Throttling outgoing spam. SpamTracker complements previous
proposals that have suggests throttling senders using schemes such
as stamps, proof-of-work, etc. One prominent postage scheme is
called “bankable postage”, whereby senders obtain stamps or to-
kens from some authority and then attach these tokens to emails [2,
41]. Other techniques for throttling spam require the sender to issue
some “proof of work”, either in CPU [5] or memory [14], although
these schemes have also been criticized because, in certain circum-
stances, they can prevent legitimate users from sending normal vol-
umes of email [21].

8. CONCLUSION
This paper presented SpamTracker, a system that classifies email
senders using a technique we call behavioral blacklisting. Rather
than classifying email senders according to their IP addresses, be-
havioral blacklisting classifies senders based on their sending pat-
terns. Behavioral blacklisting is based on the premise that many
spammers exhibit similar, stable sending patterns that can act as
“fingerprints” for spamming behavior.
SpamTracker clusters email senders based on the set of domains
that they target. SpamTracker uses these sending patterns of con-
firmed spammers to build “blacklist clusters”, each of which has
an average vector that represents a spamming fingerprint for that
cluster. SpamTracker tracks sending patterns of other senders and
computes the similarity of their sending patterns to that of a known
spam cluster as the basis for a “spam score”. Our evaluation us-
ing email logs from an email provider that hosts over 115 inde-
pendent domains shows that SpamTracker can complement existing
blacklists: it can distinguish spam from legitimate mail and also de-
tects many spammers before they are listed in any blacklist. Spam-
Tracker’s design makes it easy to replicate and distribute, and de-
ploying it requires only small modifications to the configurations
of existing mail servers. Our ongoing work involves gathering data
from a wider set of domains, improving the behavioral classifica-
tion algorithms (e.g., by using other features of email senders), and
deploying the system to allow us to evaluate it in practice.
Acknowledgments. This work was supported by an NSF CAREER
award CNS-0633974, by NSF Cybertrust awards CNS-0721581
and CNS-0716278, by NSF grant CCF-0721503, and in part by the
Georgia Tech Algorithms and Randomness Center (ARC) Think-
Tank (http://www.arc.gatech.edu/). We thank Merrick Furst,
Suresh Ramasubramanian, Criag Sprosts, and Muhammad Mukar-
ram bin Tariq for comments and help.
REFERENCES
[1] Spamhaus delisting policy, 2007. http://www.spamhaus.
org/sbl/policy.html.
[2] M. Abadi, A. Birrell, M. Burrow, F. Dabek, and T. Wobber.
Bankable Postage for Network Services. In Proc. Asian
Computing Science Conference, Dec. 2003.
[3] E. Allman, J. Callas, M. Delany, M. Libbey, J. Fenton, and
M. Thomas. DomainKeys Identified Mail (DKIM) Signatures,
May 2007. http://www.ietf.org/rfc/rfc4871.txt.
[4] D. S. Anderson, C. Fleizach, S. Savage, and G. M. Voelker.
Spamscatter: Characterizing Internet Scam Hosting
Infrastructure. In Proc. 16th USENIX Security Symposium,
Boston, MA, Aug. 2007.
[5] A. Back. Hashcash. http://www.cypherspace.org/
adam/hashcash/.
[6] A. Brodsky and D. Brodsky. A Distributed Content
Independent Method for Spam Detection. In First USENIX
Workshop on Hot Topics in Understanding Botnets
(HotBots), Cambridge, MA, Apr. 2007.
[7] D. Cheng, R. Kannan, S. Vempala, and G. Wang. A
divide-and-merge methodology for clustering. ACM
Transactions on Database Systems, 31(4):1499–1525, 2006.
[8] R. Clayton. Stopping Spam by Extrusion Detection. In First
Conference on Email and Anti-Spam (CEAS), Mountain
View, CA, July 2004.
[9] R. Clayton. Stopping Outgoing Spam by Examining
Incoming Server Logs. In Second Conference on Email and
Anti-Spam (CEAS), Stanford, CA, July 2005.
[10] Cloudmark Authority Anti-Spam. http://www.
cloudmark.com/serviceproviders/authority/spam/,
2007.
[11] Commtouch Inc. 2006 Spam Trends Report: Year of the
Zombies. http://www.commtouch.com/documents/
Commtouch 2006 Spam Trends Year of the Zombies.
pdf.
[12] E. Damiani, S. de Vimercati, and P. Samarati. P2P-Based
Collaborative Spam Detection and Filtering. In 4th IEEE
Conference on P2P, 2004.
[13] Z. Duan, K. Gopalan, and X. Yuan. Behavioral
Characteristics of Spammers and Their Network Reachability
Properties. In Proc. IEEE ICC, Glasgow, Scotland, June
2007.
[14] C. Dwork and M. Naor. Pricing via Processing or Combatting
Junk Mail. In CRYPTO, Santa Barbara, CA, Aug. 1992.
[15] P. Graham. Better Bayesian Filtering. http://www.
paulgraham.com/better.html.
[16] IronPort. http://www.ironport.com/, 2007.
[17] IronPort Carrier Grade Email Security Appliance. http://
www.ironport.com/products/ironport x1000.html,
2007.
[18] R. Kannan, S. Vempala, and A. Vetta. On clusterings: good,
bad and spectral. J. of the ACM, 51(3):497–515, 2004.
[19] Kelly Jackson Higgins, Dark Reading. Botnets Battle Over
Turf. http://www.darkreading.com/document.asp?
doc id=122116, Apr. 2007.
[20] J. Kong et al. Scalable and Reliabile Collaborative Spam
Filters: Harnessing the Global Socail Email Networks. In 3rd
Annual Workshop on the Weblogging Ecosystem, 2006.
[21] B. Laurie and R. Clayton. Proof-of-Work Proves Not to
Work. In Third Annual Workshop on Economics and
Information Security (WEIS), Minneapolis, MN, May 2004.
[22] F. Li and M.-H. Hseih. An Empirical Study of Clustering
Behavior of Spammers and Group-based Anti-Spam
Strategies. In 3rd Conference on Email and Anti-Spam
(CEAS), Mountain View, CA, July 2006.
[23] MailAvenger, 2007. http://www.mailavenger.org/.
[24] Mail Abuse Prevention System (MAPS). http://www.
mail-abuse.com/.
[25] Messaging Anti-Abuse Working Group. MAAWG Issues
First Global Email Spam Report. http://www.
prnewswire.com/cgi-bin/stories.pl?ACCT=104&
STORY=/www/story/03-08-2006/0004316196, Mar.
2006.
[26] PandaLabs Blog. Zunker Spamming Bot Front-end Analysis.
http://blogs.pandasoftware.com/blogs/
pandalabs/archive/2007/05/08/Zunker.aspx, May
2007.
[27] V. Prakash. Vipul’s Razor. http://razor.sourceforge.
net/, 2007.
[28] Pyzor. http://pyzor.sourceforge.net/, 2007.
[29] A. Ramachandran, D. Dagon, and N. Feamster. Can DNSBLs
Keep Up with Bots? In 3rd Conference on Email and
Anti-Spam (CEAS), Mountain View, CA, July 2006.
[30] A. Ramachandran and N. Feamster. Understanding the
Network-Level Behavior of Spammers. In Proc. ACM
SIGCOMM, Pisa, Italy, Aug. 2006.
[31] Secure Computing. http://www.securecomputing.com/,
2007.
[32] Secure Computing IronMail. http://www.
securecomputing.com/index.cfm?skey=1612, 2007.
[33] A. Sinclair and M. Jerrum. Approximate counting, uniform
generation and rapidly mixing markov chains. Information
and Computation, 82:93–133, 1989.
[34] Spam and Open-Relay Blocking System (SORBS). http://
www.sorbs.net/.
[35] SpamAssassin, 2007. http://www.spamassassin.org/.
[36] SpamCop. http://www.spamcop.net/.
[37] Spamhaus, 2007. http://www.spamhaus.org/.
[38] SpamHINTS. http://www.spamhints.org/, 2007.
[39] Realtime uri blacklist. http://www.uribl.com/.
[40] P. Vixie. Distributed Checksum Clearinghouse. http://
www.rhyolite.com/anti-spam/dcc/, 2007.
[41] M. Walfish, J. Zamfirescu, H. Balakrishnan, and D. Karger.
Distributed quota enforcement for spam control. In Proc. 3rd
Symposium on Networked Systems Design and
Implementation (NSDI), San Jose, CA, May 2006.
[42] M. Wong and W. Schlitt. Sender Policy Framework (SPF) for
Authorizing Use of Domains in E-Mail, Apr. 2006. RFC
4408.