SRUTI '06 Abstract
Pp. 5560 of the Proceedings
Leveraging Good Intentions to Reduce Unwanted Network Traffic
Marianne Shaw
marianne.shaw@gmail.com
Abstract:
We present a solution to reduce unwanted network traffic by enabling
either side of a conversation to summarily terminate the conversation
without the other endpoint's cooperation. Our work is motivated
by the observation that many compromised endhosts on the network are
well-intentioned but easily compromised; these machines are often
compromised and their resources used to attack others. We argue that
the good intentions of these endhosts can be leveraged to construct a
control plane that ensures that, even when compromised, these
well-intentioned machines only generate well-behaved traffic. This
independently enforced control plane prevents an endhost from
blatantly disregarding requests to cease traffic generation. The
solution's viability rests upon its unobtrusive deployment. No extra
mechanism is needed within the network as all enforcement is performed
at the endhosts. Hosts are not restricted in their behavior except by
the behavior demanded by their peers.
1 Introduction
Commodity PCs and broadband have enabled huge numbers of users to
connect to the Internet. Once connected to the Internet,
user-administered machines are bombarded with attacks aimed at gaining
control of their physical resources. Compromised machines are used to
propagate worms and viruses, participate in DDoS attacks, provide
services such as spam relays or IRC servers, and be part of organized
botnets. Users do not want their machines to be compromised and used
for malicious purposes, but they do not have the knowledge or skill to
prevent it. This work asks, can we leverage the users' non-malicious
intentions to prevent their machines from being used to generate
unwanted traffic?
This work is driven by several key observations. First, we accept
that a machine will be compromised and attempts will be made to use it
to generate malicious network traffic. Attackers use compromised
machines to amplify their ability to inflict damage; we can inhibit
their potential impact by reducing the benefits of incorporating these
machines. Next, we believe that many user-administrators do not want
their machines to be used to inflict damage, and they would be willing
to thwart such activity if they could. There is benefit to preventing
the injection of malicious traffic into the Internet, rather than
trying to deal with it once it is inside the network. We can leverage
users' good intentions to co-locate an enforcement mechanism with a
host and use it to prevent the injection of certain traffic into the
network. Finally, defining and identifying unwanted behavior is
difficult and often subjective; two hosts may not classify the same
traffic in the same way. Rather than propose a universal definition of
good or bad traffic, we seek to provide a way for traffic recipients
to request the temporary cessation of traffic that they themselves
deem undesirable.
Our solution puts control of the network packet exchange between two
hosts in the hands of both of the endpoints in such a way that each
endpoint has complete control over it. We do this by constructing an
independent control plane that is co-located with each well-meaning
host. When one endpoint requests that the other temporarily stops
sending traffic to it, the control plane prevents that second host
from disregarding the termination request; all outgoing packets are
immediately dropped at the generating host and never enter the
Internet. This conversation ``ripcord'' is necessary because a
compromised host may have an altered networking stack or operating
system [11] which ignores all (if they exist) standard
termination requests. As long as both hosts believe that the
conversation is well-behaved, there is minimal impact on the
conversation.
2 Approach
Computers connected to the Internet continue to be attacked and
compromised. User-administered machines are especially vulnerable to
compromise as they often run unpatched computer programs that allow
attackers to capitalize upon well-documented software flaws. Symantec
[23] found that several popular unpatched desktop
operating systems were compromised within 1.5 hours of being connected
to the Internet. Once compromised, the physical resources of a
machine may be used to further propagate an attack or may be
incorporated into a network of attack machines.
We seek to prevent well-intentioned machines from being compromised
and used to amplify an attacker's ability to inflict damage on the
Internet. To do this, we make it possible for a host to tell a
machine from which it is currently receiving packets to temporarily
stop sending packets, and to reasonably expect that the request will
be honored. Thus, if a compromised host is sending network attack
traffic, a receiving host can stop the incoming packets to protect
itself. While this work focuses on enabling endhosts to request the
cessation of an attack, this mechanism could be used by overloaded
hosts to temporarily delay incoming traffic.
There are a few key requirements for achieving our goals. 1) Upon
receiving unwanted network traffic, a host must be able to identify
the source of the traffic to which it can send a termination
request. Any host that voluntarily adopts our mechanism must therefore
be prevented from spoofing its packets' source address. 2) We only
honor requests to temporarily terminate an existing packet stream. We
do not allow hosts to proactively blacklist traffic that they have not
yet received; nor do we honor termination requests for packet streams
that have been inactive for a long duration. 3) Only a recipient of
unwanted network traffic can request that a packet stream be
terminated; malicious hosts should not be able to use this mechanism
to force a well-intentioned host into silence. 4) Our enforcement
mechanism must be voluntarily adopted by endhosts. We cannot rely upon
the introduction of new mechanism within the network itself. We cannot
impose undue restrictions upon the network services used by the host
simply to accommodate our mechanism. 5) Upon receiving a termination
request, we must be able to terminate a packet stream without the
receiving machine's cooperation. Once a well-intentioned machine has
been compromised, rootkits [9] can be used to gain
superuser privileges on a machine; the machine's networking stack and
OS can then be modified, replaced, or subverted [11].
Typically, a machine's OS and networking stack enforce
well-established ``good behavior;'' once these modules are
compromised, a machine is able to blatantly disregard all standard
networking conventions.
We leverage the good intentions of non-malicious users to co-locate an
enforcement mechanism with each host. The mechanism itself must be
independent, not network-addressable, and able to interpose on all
traffic in to and out of the host. Provided it can meet all of these
specifications, the mechanism itself can be implemented in either
hardware, software, or a combination of both.
Although each mechanism only enforces our requirements for an
individual host, in aggregate these mechanisms create an independent
control plane. This plane ensures that all traffic that it allows to
enter the Internet can be summarily terminated at a recipient's
request. Recipients of unwanted network traffic now have a course of
action by which they can protect themselves without requiring the
cooperation of their ISP.
3 Design
Our control-plane enforcement mechanisms must ensure that potential
victims can accurately identify their attackers from the offending
packet stream, determine the validity of requests to temporarily stop
an existing packet stream, and enforce valid network traffic
termination requests without endhost cooperation. A combination of
control plane signalling and our enforcement mechanism make this
possible.
To leverage users' good intentions, our approach must be able to
provide significant benefits without introducing a system
administration burden on the users. Once a user allows us to interpose
on all network traffic in to and out of their machine, all necessary
information should be gleaned from the traffic that we observe. From
the stream of packets, we must be able to identify the co-located
host's unique identifier, the start and end of a conversation between
two hosts, and a termination request.
Unique Identifier Ideally, the Internet would provide each host
with a unique non-forgeable identifier that could be used to provide
accountability for actions taken by a networked host. In practice,
hosts have the ability to transmit arbitrary network packets; they can
assume another host's identity by transmitting packets with spoofed
source addresses. Additionally, rather than having a single assigned
static IP address, many hosts dynamically acquire their IP address for
a short period of time.
Accountability is a necessary element of our solution. A recipient of
unwanted traffic must be able to identify and contact the host that
sent the traffic; at the same time, that host should not be penalized
for spoofed packets sent by other machines. We must ensure that
well-intentioned hosts cannot send packets with spoofed source
addresses, but we must also monitor the packets that our host did send
to deny accountability for others' actions.
A host's unique identifier is therefore the source IP address that it
used to send a stream of packets during a specific time period. This
approach, while not perfect, is sufficient for our purposes because we
only honor valid termination requests for active traffic streams.
Our enforcement mechanism can determine a host's unique identifier
from the consistent source address of the packets that it sends. We
can leverage events that signify an expected change in IP address to
recognize valid IP address changes. For example, hosts that
dynamically acquire IP addresses tend to exhibit lulls in their
network activity before they acquire a new IP address. Alternatively,
if the control mechanism is directly connected to the host's network
card, it can detect when a card is reset by the link going down. These
events occur over a period of seconds. In contrast, many network
attacks rapidly send packets with quickly changing spoofed source
addresses; our mechanism should characterize these as spoofed packets
and drop them.
We can prevent our host from being penalized for spoofed packets sent
by other machines by tracking the packets that were actually sent by
the host. Rather than log each individual packet transmitted, we can
track the fact that we sent packets associated with a particular
network conversation (defined below) during a certain time frame. When
presented with a termination request for a packet that the host did
not send, the enforcement mechanism simply discards the
request. Because termination requests are only honored for active
packet streams, the amount of state required is bounded by the number
of currently active streams.
Defining a network conversation A network conversation defines
both the criteria and the granularity that we use to track sequences
of network packets; it dictates which packets will be dropped when a
termination request is received. Hosts receiving unwanted traffic
must weigh the cost of receiving those incoming packets against the
cost of terminating the network conversation.
We can provide the ability for a host to summarily terminate a
conversation for many different definitions of a network conversation
provided we are able to uniquely identify the principals of a
conversation from each packet sent by the host, identify the start and
stop of a conversation by observing the packet stream, and identify
the termination request associated with a conversation.
Conversation principals Traditionally, IP source and destination
addresses have been the basis for identifying network
conversations. Additional properties such as IP protocol and source
and destination ports have been used to refine these principals. We
can extend this set to include more coarse grained principals. IP
prefixes could be used instead of IP addresses; thus, our enforcement
mechanism can honor requests to drop all UDP port 666 traffic destined
for 10.10.10.*. Alternatively, a host can request that it no longer be
sent any TCP traffic.
Conversation start/stop The enforcement mechanism must know
exactly what indicates that a conversation is active and inactive.
Ideally, we can identify the start and stop of a conversation simply
by observing the contents of network packets and maintaining internal
state. For example, TCP uses explicit start, stop, and termination
sequences for maintaining connections. We can use this protocol
signalling to restrict and terminate wayward conversations; a
prototype for TCP is outlined in Section 4.
However, for many conversations there is no explicit signalling
indicating the conversation delimiters, and we must infer the start
and stop of the conversation by observing patterns of network
activity. Correctly inferring these endpoints can be difficult;
although we can use the existence of network traffic between two
principals to recognize that a conversation is active, for many
long-lived conversations we cannot use the absence of network traffic
to determine that a conversation has been stopped.
Termination requests Hosts require an explicit signalling
mechanism for terminating a network conversation. In addition to
indicating which network conversation is being terminated, these
requests must either indicate or imply the amount of time during which
packets must not be sent; we do not allow network conversations to be
terminated indefinitely.
Certain protocols may have existing support for terminating a
conversation; for example, TCP uses RST packets to reset a
conversation. However, if we must infer active conversations based
upon the existence of network traffic between two principals, it is
unlikely that there will already be an existing explicit signal for
terminating the conversation. To accommodate these ad-hoc definitions
of conversations and enable hosts to reliably terminate them, it may
be necessary to provide a new signalling mechanism.
Termination requests must demonstrate that the request is being sent
by the recipient of the unwanted network traffic; spoofed termination
requests should be discarded. The requesting host can be authenticated
through the exchange of a large random nonce with the enforcement
mechanism. If the nonce cannot be overlaid on top of a network
conversation's existing protocol, then an explicit authenticating
nonce exchange may be required. Once successfully exchanged, the nonce
can be injected into a termination request to establish its
authenticity.
To convince well-intentioned users to allow our enforcement mechanisms
to interpose on their machine's network traffic, we must be
unobtrusive yet effective at preventing their machines from attacking
other hosts.
The enforcement mechanism cannot be bypassed or subverted by
attackers The enforcement mechanism must interpose on all traffic in
to and out of a machine, and it must remain completely isolated and
independent from that machine. If the enforcement mechanism is not
independent, when the host machine is compromised the attacker can
simply ``turn off'' all packet-restricting components. Incapacitating
the enforcement mechanism should require physical access to the host
machine to prevent it being silently disabled by anonymous attackers.
The enforcement mechanism must actively participate in each
conversation that it may need to forcibly terminate. This work aims to
reduce attack traffic that is generated by a compromised host; in this
scenario, both sides of the enforcement mechanism are controlled
by the attacker. If the enforcement mechanism does not inject itself
into a packet stream, the compromised machine can collude with an
external attacker to prevent a conversation's termination.
Actively injecting a nonce into a packet stream enables the
enforcement mechanism to independently authenticate an endpoint. Only
hosts directly on the path taken by outgoing network packets will be
able to reliably establish, maintain, or terminate a conversation.
The enforcement mechanism cannot be undermined by replaying a
previous conversation through the mechanism This is especially
important as many hosts acquire their IP addresses dynamically; an
attacker could try to replay a previous conversation to inflict damage
on the host now allocated a specific IP address. Therefore, the
enforcement mechanism must require proof of ``liveness'' for all
conversations flowing through it. The nonces used to authenticate
endpoints should be randomly generated at the time they are needed.
The enforcement mechanism can be deployed incrementally by end
users and removed as needed, which should be extremely rare. The
enforcement mechanism must be effective at reducing unwanted network
traffic as it is incrementally deployed. Not all user-administered
machines are going to immediately install a mechanism that prevents
their machines from being used to attack others; indeed, not all
users will want to install such a mechanism. The enforcement
mechanism must not rely upon upgraded hardware within the network or
widespread deployment and adoption of new protocols. By co-locating
the enforcement mechanism with the hosts that they are potentially
restricting, our solution can be deployed by individual users without
requiring ambitious network hardware or software upgrades.
4 TCP Prototype
We describe a prototype implementation of our solution for
TCP. Because TCP is a connection-oriented protocol, we were able to
use its existing characteristics to develop a prototype that is
virtually invisible to endhosts. Our enforcement mechanism executes
on a separate physical machine whose sole purpose is to act as a
gateway between our user-administered host and the larger network.
All traffic to and from the host must pass through our enforcement
mechanism over the dedicated Ethernet connection.
The system ``learns'' the host's IP address by observing the source IP
address in all outgoing network packets. If the network link goes down
or if there is a sustained period of network inactivity, the system
re-learns the IP address when outgoing packets are observed. All
outgoing packets using a different source IP address are dropped.
We define our network conversation to be the connection established
between two (IP:port) pairs using TCP's three-way handshake
protocol. We leverage TCP's handshake protocol to determine the start
of a network conversation. TCP also contains two distinct techniques
for closing a connection: a FIN-ACK sequence initiated by each half of
the connection, and a RST packet sent by either side of the
connection. As with TCP's connection establishment, we simply leverage
this explicit signalling to track the end of a conversation.
As long as neither host is compromised or misbehaving, TCP's built-in
control signalling ensures that hosts can terminate any undesired
connection. The true merit of our enforcement mechanism is observed
when one of the hosts is ignoring the TCP termination messages that it
receives.
Imagine that a remote host establishes a TCP connection with our local
host, and the local host starts flooding the remote host with network
packets. The remote host may send a RST packet to stop the packet
flood, but the local host may simply ignore the RST packet and
continue to send high rates of unwanted packets. Our enforcement
mechanism monitors each established connection to prevent this type of
scenario. Once it observes a valid incoming RST packet, the
enforcement mechanism drops all outgoing network packets associated
with this connection.
In its efforts to restrict unwanted outgoing traffic, the enforcement
mechanism must be careful not to allow spoofed RST packets to cause it
to incorrectly terminate TCP connections. Additionally, it must
prevent a compromised host and a colluding remote attacker from
spoofing a connection establishment sequence and using it to attack a
third network host. TCP uses the exchange of sequence numbers to
provide reasonably good authentication of each endpoint during
connection establishment and teardown. This approach, however, relies
upon the belief that at least one of the participating hosts is
well-behaved and trustworthy. Because both sides of our enforcement
mechanism are potentially compromised, it cannot rely upon the
validity of TCP's authentication.
Our enforcement mechanism must provide its own endpoint
authentication; it does this by adding a random 32-bit nonce to the
initial sequence number (ISN) provided by each host during connection
establishment. By adding this random value to each host's sequence
number, the enforcement mechanism authenticates each endpoint when the
modified sequence number is returned. This is the same authentication
technique used by standard TCP, but when used by the enforcement
mechanism it ensures that two untrusted, colluding hosts cannot
subvert the enforcement mechanism using pre-established ISNs. The
random nonces are individually generated for each connection
establishment seen by the enforcement mechanism; thus the nonce
provides the ``liveness'' property necessary for thwarting replay
attacks.
Adding the authenticating nonce to TCP's sequence number requires the
enforcement layer to continue modifying all subsequent packets'
sequence numbers. It must add the nonce to all outgoing packets'
sequence numbers and remove the nonce from all incoming packets'
sequence numbers. The enforcement mechanism must also recalculate the
checksum for each modified packet; the checksum does not need to be
completely recalculated but can simply be updated by the difference
between the old and new fields.
The enforcement mechanism maintains per-connection state to track the
status of each TCP connection. Our implementation required 108 bytes
of connection state for each active connection. Because our mechanism
enforces conversation termination for a single host, the number of
active connections that we are monitoring should remain small, as will
our overall storage requirements.
5 Related Work
A diverse set of techniques and mechanisms have been proposed to
address the widespread, damaging nature of modern Internet
attacks.
A variety of projects have attempted to characterize network traffic;
these characterizations can then be used to filter or identify
unwanted network traffic. Network connectivity patterns have been
used to characterize both normal and abnormal network traffic
( [3], [25], [22], [24])
including the propagation patterns of individual worms
( [15], [14], [28].) Studies have
quantified denial-of-service [16] activity and
spyware [19] seen on the Internet. Worms signatures
( [10], [17], [21]) can be
used in the identification of traffic containing worms. Other work has
focused on the characteristics of ``normal'' traffic
( [12], [27]) for detecting or
rate-limiting anomalous behavior.
Our proposed solution is orthogonal to this work in that we require an
endhost to determine for itself whether or not a particular stream of
network traffic is unwanted. We provide a mechanism whereby the host
can request that a packet stream be halted; a host can leverage any of
these characterization techniques to decide if the stream is unwanted.
Existing research has proposed introducing new mechanism in the
network to identify, account for, and eliminate unwanted traffic. IP
Traceback [20] uses network state to identify the path
taken by unwanted DoS traffic. Pushback [8] and
AITF [2] install packet filters at routers within
the network to filter out unwanted traffic. Capability-based
networks [1] use packet processing hardware at
trust boundaries to enable hosts to communicate while network attacks
occur. In the face of congestion, network hardware may selectively
mark [6] or drop packets associated with high packet rates.
In contrast with these network-based mechanisms, this work proposes a
mechanism that is co-located with a host to prevent unwanted network
traffic from being injected into the Internet. We deploy our
enforcement mechanisms at potentially malicious traffic sources so
that we can drop unwanted traffic before it can impact other
hosts. This principle is similarly embraced by network ingress
filtering [5], reverse firewalls [13], and IP
throttling [26]. Although we allow hosts to push
cessation requests upstream like Pushback [8] and
AITF [2], we depend upon users' willingness to have
mechanism co-located with their hosts to eliminate their need for
increased mechanism in the network. By leveraging hardware mechanism
at the source of network traffic, our solution can be incrementally
deployed at the endhosts. No large-scale network hardware or software
upgrades are required before benefits can begin to accrue.
Unlike many source-limiting approaches, our solution does not merely
enforce a well-established definition of good behavior, such as
limiting the rates of outgoing connections, packets, or source IP
addresses. Our work leverages techniques that use feedback mechanisms
to indicate when a host is behaving poorly. RED [6] and ECN
Nonce [4] use network mechanisms to inform a host that
there is network congestion; and TCP uses packet drops to scale back
its transmission rate. Our work differs from these approaches in that
we only rely upon endhosts to provide negative feedback in the form of
requests to terminate malicious conversations.
Finally, a key property of our solution is the independence of our
enforcement mechanism. Our enforcement mechanism assumes that it is
surrounded by untrustworthy, malicious entities that will try to
subvert or disable it. Therefore, the enforcement mechanism must be
active in its efforts to prevent malicious traffic. This is in direct
contrast to many firewalls [7] and intrusion
detection systems [18] which assume that at least one side of
the enforcement mechanism is trustworthy.
We have argued that we can leverage the good intentions of users to
reduce unwanted traffic on the Internet. User-administrated machines
are frequently vulnerable to compromise, and once compromised their
physical resources can be used to attack other hosts. By co-locating
an enforcement mechanism with these well-intentioned hosts, recipients
of unwanted traffic can summarily terminate streams of incoming
packets from these hosts. If a host has been compromised and is
attacking other machines, the victims have the ability to, at least
temporarily, terminate the attack.
- 1
-
Tom Anderson, Timothy Roscoe, and David Wetherall.
Preventing internet denial-of-service with capabilities.
SIGCOMM Comput. Commun. Rev., 34(1):39-44, 2004.
- 2
-
Katerina J. Argyraki and David R. Cheriton.
Active internet traffic filtering: Real-time response to denial of
service attacks.
CoRR, cs.NI/0309054, 2003.
- 3
-
Daniel R. Ellis, John G. Aiken, Kira S. Attwood, and Scott D. Tenaglia.
A behavioral approach to worm detection.
In WORM '04: Proceedings of the 2004 ACM workshop on Rapid
malcode, pages 43-53, New York, NY, USA, 2004. ACM Press.
- 4
-
David Ely, Neil Spring, David Wetherall, Stefan Savage, and Tom Anderson.
Robust congestion signaling.
In ICNP, 2001.
- 5
-
Paul Ferguson and Daniel Senie.
Network Ingress Filtering: Defeating Denial of Service Attacks which
employ IP Source Address Spoofing RFC 2267.
IETF RFC Publication, January 1998.
- 6
-
Sally Floyd and Van Jacobson.
Random early detection gateways for congestion avoidance.
IEEEACM Transactions on Networking, 1(4):397-413, 1993.
- 7
-
Michael B. Greenwald, Sandeep K. Singhal, Jonathan R. Stone, and David R.
Cheriton.
Designing an academic firewall: Policy, practice, and experience with
surf.
In SNDSS '96: Proceedings of the 1996 Symposium on Network and
Distributed System Security (SNDSS '96), page 79, Washington, DC, USA, 1996.
IEEE Computer Society.
- 8
-
John Ioannidis and Steven M. Bellovin.
Implementing pushback: Router-based defense against ddos attacks.
In NDSS, 2002.
- 9
-
Nick L. Petroni Jr., Timothy Fraser, Jesus Molina, and William A. Arbaugh.
Copilot - a coprocessor-based kernel runtime integrity monitor.
In USENIX Security Symposium, pages 179-194, 2004.
- 10
-
Hyang-Ah Kim and Brad Karp.
Autograph: Toward automated, distributed worm signature detection.
In USENIX Security Symposium, pages 271-286, 2004.
- 11
-
Samuel T. King, Peter M. Chen, Yi-Min Wang, Chad Verbowski, Helen J. Wang, and
Jacob R. Lorch.
Subvirt: Implementing malware with virtual machines.
In SP '06: Proceedings of the 2006 IEEE Symposium on Security
and Privacy, page 65, Washington, DC, USA, 1997. IEEE Computer Society.
- 12
-
John McHugh and Carrie Gates.
Locality: a new paradigm for thinking about normal behavior and
outsider threat.
In NSPW '03: Proceedings of the 2003 workshop on New security
paradigms, pages 3-10, New York, NY, USA, 2003. ACM Press.
- 13
-
Jelena Mirkovic, Gregory Prier, and Peter L. Reiher.
Attacking ddos at the source.
In ICNP '02: Proceedings of the 10th IEEE International
Conference on Network Protocols, pages 312-321, Washington, DC, USA, 2002.
IEEE Computer Society.
- 14
-
D. Moore, V. Paxson, S. Savage, C. Shannon, S. Staniford, and N. Weaver.
The spread of the sapphire/slammer worm.
Technical report, CAIDA, ICSI, Silicon Defense, UC Berkeley EECS and
UC San Diego CSE, 2003.
- 15
-
D. Moore, C. Shannon, and J. Brown.
Code-Red: a case study on the spread and victims of an Internet
worm.
In Proc. of Internet Measurement Workshop 2002, Nov 2002.
- 16
-
David Moore, Geoffrey M. Voelker, and Stefan Savage.
Inferring Internet denial-of-service activity.
In USENIX Security Symposium, 2001.
- 17
-
James Newsome, Brad Karp, and Dawn Song.
Polygraph: Automatically generating signatures for polymorphic worms.
In IEEE Symposium on Security and Privacy, 2005.
- 18
-
Vern Paxson.
Bro: a system for detecting network intruders in real-time.
Computer Networks (Amsterdam, Netherlands: 1999),
31(23-24):2435-2463, 1999.
- 19
-
Stefan Saroiu, Steven D. Gribble, and Henry M. Levy.
Measurement and analysis of spyware in a university environment.
In NSDI, pages 141-153, 2004.
- 20
-
Stefan Savage, David Wetherall, Anna Karlin, and Tom Anderson.
Practical network support for ip traceback.
In SIGCOMM '00: Proceedings of the conference on Applications,
Technologies, Architectures, and Protocols for Computer Communication, pages
295-306, New York, NY, USA, 2000. ACM Press.
- 21
-
Sumeet Singh, Cristian Estan, George Varghese, and Stefan Savage.
Automated worm fingerprinting.
In OSDI, pages 45-60, 2004.
- 22
-
S. Staniford-Chen et al.
GrIDS--A graph based intrusion detection system for large
networks.
In Proceedings of the 19th National Information Systems Security
Conference, volume 1, pages 361-370, October 1996.
- 23
-
Symantec.
Symantec internet security threat report, volume ix, Mar 2006.
- 24
-
Godfrey Tan, Massimiliano Poletto, John Guttag, and Frans Kaashoek.
Role Classification of Hosts within Enterprise Networks Based on
Connection Patterns.
In The USENIX Annual Technical Conference 2003, San Antonio,
TX, June 2003.
- 25
-
T. Toth and C. Kruegel.
Connection-history based anomaly detection, 2002.
- 26
-
J. Twycross and M.M. Williamson.
Implementing and testing a virus throttle.
In USENIX Security Symposium, 2003.
- 27
-
Matthew M. Williamson.
Throttling viruses: Restricting propagation to defeat malicious
mobile code.
In ACSAC '02: Proceedings of the 18th Annual Computer Security
Applications Conference, page 61, Washington, DC, USA, 2002. IEEE Computer
Society.
- 28
-
Cynthia Wong, Stan Bielski, Jonathan M. McCune, and Chenxi Wang.
A study of mass-mailing worms.
In WORM '04: Proceedings of the 2004 ACM workshop on Rapid
malcode, pages 1-10, New York, NY, USA, 2004. ACM Press.
Leveraging Good Intentions to Reduce Unwanted Network Traffic
This document was generated using the
LaTeX2HTML translator Version 2002-2-1 (1.70)
Copyright © 1993, 1994, 1995, 1996,
Nikos Drakos,
Computer Based Learning Unit, University of Leeds.
Copyright © 1997, 1998, 1999,
Ross Moore,
Mathematics Department, Macquarie University, Sydney.
The command line arguments were:
latex2html -split 0 -show_section_numbers -local_icons sruti2006.tex
The translation was initiated by Marianne Shaw on 2006-06-08
Marianne Shaw
2006-06-08
|