Truth in Advertising: Lightweight Verification of Route Integrity

Truth in Advertising: Lightweight Verification of Route Integrity
Truth In Advertising:
Lightweight Verification of Route Integrity
Edmund L. Wong, Praveen Balasubramanian,
Lorenzo Alvisi, Mohamed G. Gouda, Vitaly Shmatikov
Dept. of Computer Sciences, The University of Texas at Austin
Austin, TX, USA
{elwong,praveenb,lorenzo,gouda,shmat}@cs.utexas.edu
Abstract
We design and evaluate a lightweight route verification mechanism that enables a router to discover route failures and
inconsistencies between advertised Internet routes and actual paths taken by the data packets. Our mechanism is
accurate, incrementally deployable, and secure against malicious intermediary routers. By carefully avoiding any cryptographic operations in the data path, our prototype implementation achieves the overhead of less than 1% on a 1
Gbps link, demonstrating that our method is suitable even
for high-performance networks.
Categories and Subject Descriptors:
C.2.2 [Computer-Communication Networks] Network
Protocols
General Terms: Security
Keywords: Interdomain routing, security, BGP
1.
INTRODUCTION
One of the fundamental features of the Internet is the autonomy of the individual domains which comprise it. Each
domain, or autonomous system (AS), decides independently
how to select the routes and how to forward the data traffic
it receives from its customers and peer domains. The control
plane of today’s Internet, i.e., the mechanism for establishing inter-domain routes, is based on the Border Gateway
Protocol (BGP). Each BGP route is a chain of autonomous
systems, leading to a particular set of destinations (identified
by an IP address prefix). BGP is a path vector protocol: the
routers of each autonomous system keep in their databases
the complete AS-level route leading to each reachable prefix. The autonomous system advertises chosen routes to its
peer ASes. The advertised routes need not be the cheapest,
fastest or shortest; choosing which routes to advertise and
use is an internal decision for every AS.
Upon receiving its peers’ advertisements, an AS applies
its internal policy to select the “best” route to every des-
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.
PODC’07, August 12–15, 2007, Portland, Oregon, USA.
Copyright 2007 ACM 978-1-59593-616-5/07/0008 ...$5.00.
tination and update its route database. The policies that
govern route selection by individual ASes are completely autonomous and, in general, not necessarily based on just the
next hop and/or route length. BGP route advertisements
are specifically designed to include the entire route. The reason for this is to enable selection policies that involve “deep”
route inspection beyond the next hop when deciding which
route to use. Route selection policies may take into account
service-level agreements, business contracts, previously observed performance, reputation, or any other factors. At the
network level, however, all of these considerations must be
translated into BGP policies that choose a particular route
simply because certain ASes are present or absent in it.
Route selection policies are meaningful only if the advertisements on which they are based are truthful, i.e., if ASes
forward traffic in the data plane using the routes they advertised in the control plane. Control-plane verification of
AS-level routes may prevent malicious ASes from advertising unauthorized routes (see Section 2), but they can still
deviate arbitrarily from stated routes by forwarding an AS’s
data down routes that the latter would not have selected.
In short, there is no truth in advertising. Inconsistencies between the control plane and the data plane can arise
not just because of BGP anomalies [15], router misconfiguration [14], or security compromises. Autonomous systems
may also have economic incentives to misroute data traffic. Different ASes charge different fees for carrying their
peers’ traffic. There may be a substantial financial benefit for an AS to advertise a fast route, which is more likely
to be selected by its peers and provide more revenue, but
forward data packets using a different, cheaper path. Because route selection depends on the assumption that the
advertised route is the one that will actually be used, such
inconsistencies can not only nullify the route selection policies used by a particular AS, but also have a negative impact
on stability and performance of the Internet.
AS-level traceroute experiments indicate a discrepancy
between the control and data planes for at least 8% of Internet routes [15]. This is only the lower bound on the true
scale of the problem, because none of the existing route
verification methods are secure against misbehaving ASes
determined to cover their tracks. For example, traceroutebased verification can be easily evaded by a malicious AS
who simply forwards traceroute packets down the “correct”
advertised route, while diverting data packets.
Our contributions. We present a lightweight protocol
that enables a router, acting as the verifier, to verify that
its data traffic follows a certain route through the Internet.
Our protocol does this by verifying the presence of a given
autonomous system (the prover ) and its predecessor on the
data path. By chaining these verifications, the verifier can
localize faults to a particular segment of the route.
Our main objective is high-performance verification. This
is essential for verifying Internet routes, because the provers
may be heavily loaded AS-interconnect routers, which potentially process billions of packets per second. We carefully design our verification protocol so that it (i) completely
avoids cryptographic operations in the data path, and (ii)
does not require the prover to maintain any long-term flowspecific state. Instead, the prover merely maintains a flowindependent table to efficiently participate in multiple concurrent instances of the verification protocol.
The cost of avoiding cryptographic operations in the data
path is an offline setup phase. It is executed infrequently via
an out-of-band channel (e.g., a TLS-protected website) and
enables the prover and the verifier to share a set of (2k + 1)bit random secret tuples, where k is the security parameter.
We emphasize that the prover and the verifier do not need
to maintain a secure channel outside of this setup. Moreover, the prover does not have to manage individual, online
cryptographic state for each verifier. It is sufficient for the
prover to remember the secrets that it recently generated,
but not with whom they are shared.
To check integrity of some route which is supposed to
include the prover, the verifier randomly embeds the first
element of a shared secret tuple into some of the data packets
on the route. These are the probe packets. In the rest of
the data packets, the verifier embeds fresh random numbers.
The prover checks for these tags in the incoming packets and
responds with the second element in the shared secret tuple
to those containing the first element. This checking need not
be done in the critical path and has a minimal impact on
router performance. The packet itself is forwarded normally.
Our verification method is secure against malicious adversaries in the data path who may be trying to divert or
drop data packets because probe packets are indistinguishable from normal data packets for everyone except the prover
and the verifier. Our protocol is completely decentralized
and assumes no more knowledge at each router about the
state of the network than provided by standard control-plane
protocols such as BGP. It is also incrementally deployable
and requires no changes to TCP/IP.
We give a security analysis of our protocol, and show
that our prototype implementation, using the Click software
router [13], introduces only 0.38 to 6.10% overhead over 1
Gbps links with almost no impact on latency.
Our focus is on the secure detection of faults in the advertised routes, whether they are caused by a discrepancy
between the control and data planes or by a failure in an
intermediary router. What to do when the verifier detects
a problem is a matter of its own BGP policy. For example,
the verifying AS may update its local route selection policy
to avoid routes that include certain ASes.
We assume that the prover cooperates in the route verification procedure. In typical Internet scenarios, a rational
AS would not want to appear in route advertisements from
other ASes if they do not use it (and pay for its services)
when the route is selected. Therefore, it has an incentive to
help verifiers ensure that if they have chosen a route containing this AS, their data traffic indeed passes through it.
We stress that we do not aim to prevent any AS from
re-routing its peers’ traffic at will as long as its data-plane
behavior is consistent with its control-plane behavior. An AS
may switch routes as long as it informs its peers by propagating the corresponding route advertisements. Upstream
ASes can then re-apply their route selection policies and
decide if they still wish to continue using the route.
Organization of the paper. We describe related work
in Section 2, our protocol in Section 3, and analytic and
experimental evaluation in Sections 4 and 5, respectively.
In Section 6, we discuss alternative implementations of our
protocol, and conclude in Section 7.
2.
RELATED WORK
Much research has been devoted to securing the control
plane of inter-domain routing by authenticating route update messages [9, 12, 23]. We focus instead on verifying the
integrity of the data plane, i.e., testing that the actual data
paths are consistent with the route advertisements.
Other methods for data-plane verification include secure
traceroute [17] and stealth probing [1]. Secure traceroute
requires each verifier-prover pair to establish a shared secret
predicate, which is used to mark certain packets as probe
packets. The prover’s responses include the next hop and are
authenticated using message authentication codes (MACs).
Stealth probing is even more expensive, requiring a complete
IPsec tunnel between the prover and the verifier in order to
hide probe packets from intermediate routers.
Both solutions are mainly intended for verifying end-toend or edge-to-edge connectivity rather than full route integrity (i.e., data packets may still be misrouted in the middle of the path). First, the prover must establish and maintain a separate cryptographic state for each verifier. A single
AS may be engaged in thousands of concurrent verifications,
requiring an expensive state management mechanism.
Second, cryptographic operations in the critical path of
the prover’s routers impose a significant throughput penalty
when the routers are fully loaded. Stealth probing requires
the router to decrypt every received packet; secure traceroute requires the router to compute a MAC for each response
to a probe packet, in addition to the possible cryptographic
operations involved in recognizing probe packets. Our preliminary tests on the current (February 2007) Click source
code [13], which incorporates the IPsec implementation used
by stealth probing, indicate a 88% drop in throughput, from
around 620 Mbps to 71 Mbps, even after increasing the MTU
to prevent fragmentation and adding headroom in the driver
to prevent expensive packet copies. It is unlikely that approaches using online cryptography can scale to inter-AS
traffic levels and be used to verify a full AS route.
Secure traceroute can also use non-cryptographic predicates on any sufficiently random packet field to identify
probe packets. To prevent a malicious intermediate router
from learning the predicate, all responses must be buffered
at the prover until the next predicate is securely established.
This neither provides an immediate feedback to the verifier
about the state of the route, nor allows fine-grained control
over the probing rate. If the predicate is updated frequently,
the prover and verifier need an online cryptographic channel,
which is problematic if the prover is a core AS.
Goldberg et al. [8] present protocols for detecting and localizing the sources of packet loss in the presence of malicious routers. They rely on cryptographic operations in
the data path, and, in the absence of performance measurements, it is not clear whether they scale to AS-level verification. The fault localization protocol of [8] can only check the
presence of a prover on the route, and it is unclear whether
it can be chained to verify the absence of unwanted ASes.
(To the best of our knowledge, the research described in [8]
was done concurrently and independently of this paper.)
Other techniques include Listen and Whisper [22], which
focuses on end-host reachability, and thus cannot detect
whether data packets have been diverted to a different AS
path as long as they reach their destination, and Fatih [16],
which requires routers to collect and share detailed statistics
for each route in order to detect routing anomalies. Collecting per-flow statistics within inter-AS routers that process
billions of packets per second does not appear feasible.
Single-packet IP traceback [20] requires routers to maintain Bloom filters with the digests of packets that traversed
them. Since AS routers cannot distinguish probes from data
traffic, they must keep track of all packets and, to support
route verification, allow verifiers secure access to their filters. Storage alone requires roughly 12 GB per minute on a
core Internet router [20]. In [19], routers mark each packet’s
16-bit fragmentation identifier field with a unique identifier.
This scheme is not secure if routers are malicious, since they
can easily mark packets with other routers’ identifiers.
3.
ROUTE VERIFICATION PROTOCOL
Our main protocol enables an AS V (the verifier ) to check,
given an advertised route containing some AS P (the prover ),
whether a particular data packet flow indeed traverses P
(more precisely, some border router belonging to P ) and the
AS that precedes P according to the route advertisement.
In general, our protocol allows any AS in the network to
act as the verifier, but specific implementations may impose
the restriction that only a single verifier can be verifying a
given packet flow at any time. In this case, an edge AS would
typically act as the verifier. For example, our prototype implementation, described in detail in Section 5, breaks up
secret tags into pieces and embeds these pieces into consecutive packets’ IP fragmentation identifier fields. Since each
IP packet header contains only one such field, only one verifier can use any given packet for route verification. Although
we expect that edge ASes will typically act as verifiers, an
intermediary AS can act as a verifier as long as the packets
used are not already tagged by some upstream AS. We also
discuss an alternative implementation method in Section 6
that supports multiple verifiers through IP options at the
expense of increased processing overhead.
For the purposes of this paper, we omit the details of the
route advertisement protocol (e.g., standard BGP can be
used), and assume that the lifetime of the route is much
longer than that of individual connections using that route.
3.1
Design principles
The route verification protocol must be secure in the presence of malicious intermediate ASes, even if the latter are
aware that route verification is taking place, know the verifier’s and prover’s algorithms, and are actively trying to
evade detection. In particular, probe packets must be indistinguishable from regular data packets. Otherwise, a malicious intermediary can forward all probe packets down the
expected route, while diverting regular packets.
Naive solutions, such as asking the destination to measure
Algorithm 1 Verifier protocol
/* data forwarding loop */
for every packet m heading to S via P do
if random(0, 1) < Pr(tag) then
if random(0, 1) < Pr(secret) then
remove some (s1 , s2 , b) from SV,P
tag m with s1
if b = 0 then
T1 = T1 ∪ {s2 }, T2 = T2 ∪ {s2 ⊕ routeid(m)}
else
T1 = T1 ∪ {s2 ⊕ routeid(m)}, T2 = T2 ∪ {s2 }
end if
else
tag m with fresh k-bit random number s̄1
end if
end if
forward m to next hop
end for
/* response handling loop */
for every response m that contains tag t do
if t ∈ T1 then
T1 = T1 \ {t}, nvalid = nvalid + 1
else if t ∈ T2 then
T2 = T2 \ {t}, ninvalid = ninvalid + 1
end if
end for
its received traffic rate, do not solve the problem when both
the expected and diverted routes lead to the same end host.
Asking the prover’s routers to measure traffic rates or other
flow characteristics and communicate them to the verifier requires the prover to potentially maintain long-term per-flow
state and does not scale to Internet-level route verification.
The same AS may appear on hundreds of thousands of
routes, and its routers may be engaged as provers in many
concurrent route verifications. Even simple operations such
as computing a cryptographic hash function on every packet
do not scale to the line speeds of modern AS-interconnect
routers. To minimize the prover’s overhead, our protocol
does not require any online cryptographic operations on data
packets. The prover and the verifier need not maintain pairwise secure channels and/or cryptographic associations.
Our use of cryptography is limited to an offline setup
phase (see Section 3.4). In Section 6, we describe an alternative that uses two cryptographic hash functions during
route verification without online pairwise secure channels.
3.2
Verifying presence of an AS
We now describe our main protocol which allows the verifier to check the presence of a certain prover on a given
data path. We extend this protocol to additionally verify
the identity of the preceding AS in Section 3.3.
Consider an abstract route shown in Fig. 1. A border
router in AS V (the verifier ) is using this route to send traffic
to an edge network S. E and P are intermediate ASes on
the route. V wants to verify that P (the prover ) is indeed
present on the data path. We assume that P is cooperating,
but E may be malicious. P may be concurrently engaged in
other instances of the protocol with other verifiers V0 .
In Section 3.4, we describe an offline setup phase, after
which V and P share a set SV,P of secret tuples (s1 , s2 , b)
R1
V!
P
S
E
V
R2
Figure 1: Routes between V and S. The advertised route is E → R1 → P → S.
Algorithm 2 Prover protocol
/* data forwarding loop */
for any incoming packet m do
if m contains tag t and return address V then
Q = Q ∪ {(t, routeid(m), V )}
end if
forward m to next hop
end for
/* tag handling loop */
for (t, routeid(m), V ) in queue Q do
Q = Q \ {(t, routeid(m), V )}
if ∃(s1 , s2 , b) ∈ S such that t = s1 then
if b = 0 then
generate anonymous response m̂ ≡ s2
else
generate anonymous response m̂ ≡ s2 ⊕ routeid(m)
end if
send m̂ to V after random delay µ
end if
end for
consisting of two random k-bit secrets s1 and s2 and one
random bit b. All of P ’s border routers store the set SP ,
which contains all secret tuples known to P . This is a single
set, not partitioned by verifier.
Once verification begins, V randomly “tags” some of the
data packets with the first element s1 from any secret tuple
(s1 , s2 , b) ∈ SV,P , and the rest with a fresh, random k-bit
tag s̄1 . Each secret tuple from SV,P is used at most once.
Along with each tag, V also embeds a return address that P
can use to contact V . Note that only V and P can tell the
difference between s1 and s̄1 . We will refer to the packets
containing secret tags as probe packets.
When a tagged packet m arrives at one of P ’s border
routers, it extracts the tag t from m. Suppose ∃(s1 , s2 , b) ∈
SP such that t = s1 . If b = 0, then P replies with s2 . If
b = 1, P replies with s2 ⊕ routeid(m), where routeid(m) is
a route identifier based on m’s source and destination ASes
or address prefixes, padded with 0’s to k bits. The reply is
sent to the return address included with the tag.
A malicious intermediary may attempt to selectively drop
replies originating from certain provers, thus “framing” them
as the source of packet loss on the route. To prevent this,
replies are anonymized by setting their source addresses to a
constant pre-defined value regardless of the prover’s IP address. Even though a malicious intermediary E cannot tell
which reply corresponds to which tag, E may attempt to
exploit timing information to determine which downstream
AS is sending the replies, e.g., by measuring the time elapsed
between the first tagged packet in the flow and the first observed response. To prevent this, the prover may delay each
reply for a random duration, proportional to the number of
hops between the prover and the destination AS.
Regardless of whether the tag is a shared secret, the data
packet itself is routed normally by the prover. With the exception of the initial tag extraction, tag analysis can be done
outside the critical packet forwarding path. If the prover’s
router is under a heavy load, it can continue forwarding
packets, queuing extracted tags for later processing.
V keeps a list of valid replies which initially contains, for
every secret tuple (s1 , s2 , b) used on packet m, s2 if b = 0
and s2 ⊕ routeid(m) if b = 1. In addition, V also keeps a list
of invalid replies, which initially contains s2 ⊕ routeid(m)
if b = 0 and s2 if b = 1. V records P ’s valid and invalid
replies. As soon as a valid or invalid reply is received, it is
counted and removed from the appropriate set (if k is sufficiently large, even a single invalid reply indicates a route
failure; see Section 4.2). Any other reply, i.e., a reply containing a number which is neither s2 , nor s2 ⊕ routeid(m)
for some (s2 , m), is simply discarded. We describe a general
statistical method for using the number of valid and invalid
replies to determine whether a route is faulty in Section 4.
To further lighten their load, P ’s routers can check only a
random fraction of the packets for tags and forward the rest
unchecked. This has the effect of increasing the drop rate
of the route as observed by the verifier since some of the
normally routed packets will not produce a valid reply, and
therefore will appear to have been dropped and/or diverted.
We discuss the implications of this in Section 4.
V may opt to tag only some of the data packets along the
route with random tags. This decreases both the verifier’s
and the prover’s overheads, but a malicious intermediary
can divert or drop the untagged packets with impunity since
they are easily distinguishable from probe packets.
3.3
Verifying predecessor AS
We now describe an extension to the basic protocol of Section 3.2 that enables the verifier to check whether the AS
preceding the prover on the data path is consistent with the
route advertisement. This is essential for chaining verifications together to check integrity of the entire route.
The main idea is to use a separate set of secret tuples
i
SR
P per each interface of P ’s border routers instead of the
same SP for all interfaces. As before, V and P share a set
of use-once (2k + 1)-bit secret tuples SV,P . During the setup
phase (see Section 3.4), V specifies, for each secret tuple it
shares with P , what the expected previous AS Ri will be.
i
The tuple is then added to the appropriate SR
P set, which is
later installed at border routers that connect P to AS Ri .
Consider again the abstract route depicted in Fig. 1,
where P is preceded by R1 in the advertised route. In this
example, for each secret tuple that V obtains from P during the setup phase, V informs P that this secret will arrive
1
from R1 . These tuples are then added only to the set SR
P ,
which is installed only at the interfaces connecting P to R1 .
If one of these secrets then later arrives to P via some other
AS R2 , the secret will be checked against a different set
2
SR
P . Since the secret does not appear in this set, the border
router will not generate a valid reply, enabling the verifier to
detect a problem with the route as described in Section 4.
This method of verification can be chained. If all provers
on the path cooperate, V can check that P1 is the predecessor of P2 , P2 is the predecessor of P3 , and so on. V simply
runs multiple instances of the protocol, maintaining a separate set of secrets and separate statistics for every prover
Pj on the route. When secrets are inserted into packets, the
appropriate sets of expected and unexpected responses are
updated. When a response comes back, V checks whether
this is an expected or unexpected response for any prover,
and, if so, updates the sets accordingly.
Completely checking a route consisting of n ASes requires
n executions of the verification protocol. Verifications can be
interleaved, i.e., consecutive data packets may contain tags
destined to different provers. If the verifier is only interested
in checking the presence of each AS, dn/2e executions are
sufficient, since a single instance of the protocol verifies the
presence of both an AS and its immediate predecessor.
By chaining verifications, the verifier can check the integrity of an entire route and determine that no additional
ASes—those not mentioned in the route advertisement—
have been inserted into the route. We describe how the
cause of faults can be localized in Section 4.
3.4
Offline setup and secret exchange
Our protocols require a setup phase, in which V securely
obtains from P a set SV,P of secret tuples (s1 , s2 , b), consisting of two random k-bit secrets s1 and s2 and one random
bit b. This is done offline, independent of route usage, and
only requires P to remember what secrets have been distributed, not who obtained them. Thus, P does not need to
manage separate cryptographic state for each verifier.
One simple method by which P can provide secrets to
potential verifiers is through a TLS-protected website, from
which an authenticated verifier Vj can securely download
a fresh random seed zj . Whenever P wants to provide a
new set of tuples to its verifiers, P publishes a long random
number r and the counter t, representing the current “generation” of shared secrets. When Vj wants to acquire a new
set of secret tuples for a particular generation t, Vj contacts
P ’s server, supplying the identity of the AS Ri that precedes
P on Vj ’s route (see Section 3.3) and the number of tuples
nj that Vj would like to generate. Applying a cryptographic
pseudo-random number generator (PRNG) to r, t and zj ,
Vj and P generate a new set of shared (2k + 1)-bit secret
tuples SVj ,P (t) corresponding to a particular generation t.
P generates the entire set SVj ,P (t), and adds it to the curi
rent set SR
P (t) corresponding to the interface connecting P
with AS Ri (specified by Vj as P ’s predecessor). Vj does not
have to generate all nj tuples at once; they can be generated
later from r, t, and zj , provided Vj remembers the state of
the PRNG. The secret sets must be completely replaced at
the prover and verifier at the beginning of a new generation.
This setup is done completely off the critical data path.
i
The next generation of shared secrets SR
P (t + 1) can be set
i
up while the current generation SR
(t)
is
still in use by P ’s
P
border routers. Verifiers are expected to check whether their
set of shared secrets is current, which can be done offline.
Even though the prover’s border routers have to store a
large set of secret tuples, we contend that the space overhead
of our protocol is feasible for modern core routers. As described in Section 5, our protocol uses a 4-way cuckoo hash
table [7]. A 2 GB 4-way cuckoo hash table can hold roughly
130 million 129-bit secret tuples (consisting of two 64-bit secrets), which is enough for approximately 4,500 ASes to send
20 secrets per minute for a day to a single incoming interface on a prover’s router. Considering that a state-of-the-art
core router such as Cisco’s CRS-1 has several gigabytes of
memory per interface card [4], our space overhead appears
practical for core Internet routers.
3.5
Route selection policies
Our focus in this paper is on developing a mechanism for
detecting and localizing route faults, not on what an AS may
do once a fault has been detected. Like other considerations
governing BGP route selection, this is an internal policy decision that is made independently by each AS. There may
be legitimate reasons for discrepancies between the control
plane and data plane, e.g., during BGP routing changes [15].
In general, each AS must decide individually how to translate the fact that a particular data path deviates from the
advertised route into a specific BGP policy.
4.
SECURITY ANALYSIS
The main purpose of our protocol is to enable verifier V
to check whether a given data route passes through a certain
AS P . In this setting, there is no conceptual difference between packets that are “naturally” dropped due to network
imperfections, those dropped by a malicious intermediate
AS E, and those diverted by E to a different route that
does not include both P and its predecessor. In all of these
cases, the packet’s actual data path is different from the
advertised route, indicating an inconsistency or route fault.
The problem of detecting route faults is effectively that
of secure network measurement. More precisely, given any
threshold δ, V must be able to tell, in the presence of a
malicious E, whether the drop rate of V ’s route to P is
above δ or not. The specific value of δ is determined by V ’s
internal policy, i.e., it is up to V to decide what drop rate
is acceptable. Our protocol does not depend on a specific δ.
We do not aim to verify complete packet integrity, which
is a much more difficult problem (arguably, packet integrity,
as opposed to route integrity, should be addressed by endto-end mechanisms such as IPsec). Similarly, we do not
aim to detect adversaries who, for example, create a copy of
each data packet, route the copy correctly and the original
packet incorrectly, or vice versa. We do not view this as
a significant limitation. It is not clear whether a malicious
intermediary AS E has any incentive to stage such an attack,
as it increases the amount of traffic E has to route. Our
protocol does ensure that if only some of the routes passing
through E are being verified, E cannot cheat by switching
tags between packets on different routes.
4.1
Tracking valid and invalid responses
V measures the connection’s drop rate δ by keeping track
of the valid and invalid responses it receives. Recall that if V
sends a packet m containing a known secret s1 , P responds
with another secret s2 , XORed with routeid(m) if b = 1.
Consequently, there are three different types of responses
that V may receive: a valid response (s2 if b = 0, s2 ⊕
routeid(m) if b = 1), an invalid response (s2 ⊕ routeid(m) if
b = 0, s2 if b = 1), or a fake response (any other value).
V ignores replies that are neither s2 , nor s2 ⊕ routeid(m).
If V counted the number of other responses, an adversary
who is not even present on the route could confuse V into
believing that the route is faulty by sending a large number
of fake responses. In our protocol, an AS must be actually
present on the data route in order to send either a valid or
invalid response to V ’s probes.
It is important, however, that V track the invalid responses it receives. Otherwise, a malicious intermediary E
could “launder” V ’s probes through a flow originating from
a different AS, one that is not performing verification. We
describe this attack in further detail in Section 4.3.
4.2
Fault detection and localization
Suppose V tags n packets with a secret tag and receives
n̂ valid responses. Each test is executed for some time τ (),
which is set to be much longer than the round-trip latency
of the connection. More specifically, if n0 represented the
number of responses generated by P in response to those of
V ’s packets that reach it, we assume that (1 − )n0 of P ’s
responses will reach V within time τ .
If the route does not include malicious routers, V should
expect to receive
(1 − )(1 − δ)2 n ≡ θn n
(1)
responses to probe packets tagged with shared secrets. Here
(1 − δ)2 is the probability that neither the probe packet, nor
P ’s response to it has been dropped in the network.
It is possible that V receives a “legitimate” invalid response if one of V ’s random tags collides with one of P ’s actual secrets s1 , shared with either V or another verifier, and
the corresponding response happens to collide with some
value in the invalid reply set. This probability is negligible
for large k, so we assume that V should expect to receive no
invalid replies. Thus, even a single invalid reply indicates
the presence of a malicious intermediary on the route.
We now describe the valid-response test in detail. Let θ =
n̂/n be the fraction of probe packets for which the verifier
receives responses. We then have two statistical hypotheses:
H1 : θ ≥ θn . The response rate to probe packets is at least
as expected.
Ha : θ < θn . The response rate is lower than expected.
We will say that the route is faulty if the verifier can reject the hypothesis H1 with high confidence. The problem of
fault detection now reduces to standard statistical hypothesis testing: what is the threshold n1 such that if verifier
V sends n probe packets and receives at most n1 responses,
V can reject hypothesis H1 ? Note, however, that attempting to minimize the false positive ratio (declaring a normal
route faulty) increases the false negative ratio (declaring a
faulty route normal), and vice versa.
To solve the problem, we use a uniformly most powerful (UMP) test [5]. For any significance level α, a UMP
test caps the false positive rate at α while guaranteeing the
lowest probability of a false negative for any statistical test
Figure 2: The tradeoff between incorrectly detecting
a fault and marking a faulty route correct for a UMP
test over the number of valid responses given 100
secret tags (n = 100), an expected response rate of
0.9 (θn = 0.9), and varying adversarial drop rates (η).
with significance α. In our setting, α is a parameter of the
verifier’s algorithm; therefore, for any given α, our UMP
test maximizes the probability of correctly detecting a fault,
while capping the probability of falsely marking a normal
route faulty to be at most α for any θn .
From the viewpoint of the verifier, the network is a “black
box.” The verifier sends a probe packet, and with some
probability a response comes back. Therefore, we model the
receipt of responses from the network as a Bernoulli process,
where the number of valid responses to n probe packets is
a random variable that follows a binomial distribution with
n trials and probability θn of success. The question “is the
route faulty?” is then equivalent to the question “is θ less
than θn ?”, which we answer with the following UMP test
for binomial distributions [5]:
For a given level of statistical significance α, find n1 such
that
Pr(n̂ ≤ n1 | θ = θn ) = α
(2)
is satisfied for a sample of size n drawn from a binomial
distribution with parameter θn .
Accept Ha (mark route as faulty) if the number of valid
responses n̂ is at most n1 (n̂ ≤ n1 ). Otherwise, if n̂ > n1 ,
accept H1 (consider the route to be normal).
In addition, given a particular UMP test of significance α
and threshold n1 , we can calculate the probability of a false
negative given a certain adversarial drop rate. If we assume
that there is an adversary who is dropping or diverting η of
the packets, the probability that our test will falsely mark
the route as being correct is
1 − F (n1 ; n, θn (1 − η))
(3)
where F (x; n, p) is the binomial cumulative density function
with parameters n and p given x successes.
In summary, for any false positive rate α, we use equation (2) to calculate a threshold n1 that guarantees this rate
and minimizes the false negative rate, which we calculate for
any adversarial drop rate using equation (3).
Fig. 2 illustrates the tradeoff for n = 100, θn = 0.9, and
various adversarial drop rates η. For a test that has signifi-
cance level α ≤ 0.01, (2) dictates that n1 = 81 (α = 0.0046).
Thus, receiving at most 81 responses occurs about 0.46% of
the time given our choice of θn and n, and we can conclude
that there is a greater than 99% chance that the route is
faulty. Given this test, if there is an adversary along the
path dropping η = 0.15 of the tags, then the probability
of correctly detecting the problem given our test (n1 = 81)
is approximately 0.883. A test with significance α = 0.05
is slightly more aggressive: receiving fewer than n1 = 84
is enough to mark the route faulty. However, this test will
detect an adversary dropping η = 0.15 of the tags with probability of approximately 0.975.
As θn decreases, either due to an increase in the network
drop rate or optimizations that throw away some of the tags
(as mentioned in Section 3.2), the number of trials necessary
to ensure statistical significance of our test increases.
Our test detects faults on the route from the verifier to
the prover or from the prover to the verifier (because of
asymmetric routing, the two routes may be different). For
example, even if the actual data path is correct, but all
replies from the prover are misrouted or dropped, the verifier
will declare the route faulty.
The verifier can test the route at discrete times, or continuously in overlapping “sliding windows.” The latter requires
the verifier to maintain a sliding list of n most recent packets, with which to compare received responses.
In addition to detecting route faults, our protocol also
helps V localize the source of the problems on the route.
Because a malicious intermediary E cannot tell which downstream prover sent a given response, it cannot selectively
drop responses from a particular prover. Therefore, whenever V observes that a route is faulty, blame for the fault
can be assigned to some AS between the last prover whose
responses arrive correctly and the first prover who does not
return the expected responses.
If chaining is done as described in Section 3.3 and there
is a single malicious AS on the route, this localizes the fault
to a pair of consecutive ASes (our protocol cannot tell the
difference between a malicious AS preceded by a good AS,
and a good AS preceded by a malicious AS who is dropping
all of its replies). In Section 4.3, we discuss the scenario
when the route contains multiple colluding malicious ASes.
4.3
Attacks
Suppose the route from verifier V to prover P contains a
malicious router E. The primary attack we are concerned
with is E dropping data packets or diverting them to a different route that does not include P . E could also potentially drop or reroute responses. We view these attacks as
equivalent in the sense that both attacks result in data not
traversing P and thus no response returning to V .
Other possible attacks include E diverting packets to a
route that includes not only P , but also additional ASes
that were not part of the route advertisement; E “framing”
a well-behaved AS by making it appear that it is misrouting
packets; E switching tags between flows; and attacks by malicious routers who are not present on the route, but try to
make the route appear faulty by generating bogus responses.
A malicious E on the route may also abuse our protocol
to stage a “reflection” attack on some victim by replacing
the verifier’s return address in tags with the victim’s address,
thereby directing the prover’s responses to the victim. However, the probe rate in our protocol is too low for this to be
an effective denial-of-service attack, and it will result in E
being detected as the cause of route fault.
Security of our protocol relies on four properties:
E cannot tell the difference between probe packets
and data packets. If E drops or reroutes a packet containing some shared secret s1 , V will not receive the expected
response. Since E cannot differentiate s1 from a random tag
s̄1 , E must route packets containing s1 down the advertised
path to avoid detection.
Because the prover’s replies are not authenticated, E may
attempt to generate a fake response to any given tag. Since
V expects responses only to probe packets, E must guess
correctly whether a given k-bit tag is a shared secret s1 and,
if so, what are the values of s2 and b in the corresponding
tuple. E’s probability of guessing correctly is |S2Vk | · 21k · 21 ,
which is negligibly small.
E cannot determine which responses contain an encoded route identifier. Although E itself cannot determine which tags are secrets, E could potentially still misroute V ’s packets. Consider a malicious intermediary AS
E which is processing two concurrent flows from ASes V
and W , respectively, both of which are supposed to traverse
some downstream AS P according to the advertised routes.
Suppose that only V is performing verification and thus only
packets in V ’s flow are tagged.
In this case, for every packet mV from V ’s flow, E may
attempt to extract the tag, insert it into a packet mW in W ’s
flow, route mW correctly and misroute mV . By forwarding
any resulting responses back to V , E can misroute V ’s traffic
yet avoid detection by V .
To prevent this, prover’s replies generated according to
the protocol of Section 3.2 depend on the value of a secret
bit b. If b = 0, the expected reply is the second element s2
of the secret tuple. Otherwise, the expected reply is s2 ⊕
routeid(mV ). Thus, for each tuple whose bit b = 1, P ’s
response to mW will be different from its response to mV .
Since E does not know the value of b for any given response
m̂, E must guess correctly whether m̂ = s2 (in which case E
must forward m̂ unmodified), or m̂ = s2 ⊕ routeid(mW ) (in
which case E must forward m̂ ⊕ routeid(mW ) ⊕ routeid(mV )
to avoid detection).
Because E knows that m̂W is either s2 or s2 ⊕routeid(mW ),
E could send both m̂W and m̂W ⊕routeid(mW )⊕routeid(mV )
to V , one of which (E does not know which) is the expected
valid response. To prevent this attack, V tracks invalid responses and declares the route faulty if it receives any.
E cannot tell which prover sent a given response. An
adversarial AS E on the route cannot tell the difference between a random tag and a shared secret when it observes a
tagged packet. E may attempt to attack the protocol by selectively dropping responses from some downstream prover
in an attempt to frame it as being faulty. However, by setting the source IP address of all replies to a default value
and randomly delaying them, our protocol prevents E from
determining which downstream prover sent a given reply.
Therefore, a single adversary E cannot selectively drop tags
destined to or responses originating from a particular downstream AS in an attempt to frame it.
If there are multiple colluding adversaries on a given route,
they can distinguish responses generated by provers who lie
between any two adversaries on the route from those generated by provers who are downstream from the last ad-
versary. For example, consider some route that contains
E1 → P → E2 , where E1 and E2 are colluding adversaries
and P represents some set of provers lying between E1 and
E2 . If a response is observed only by E1 , but not by E2 ,
they can determine that it must have originated from some
prover in P. Nevertheless, E1 and E2 still cannot tell which
prover in P produced the response, and cannot selectively
“frame” a single prover within P. Thus, V can localize the
fault to somewhere between the last successfully verified AS
and the first one for which verification failed.
E cannot change the predecessor specified by V .
Even if the data path passes through all ASes listed in the
advertised route, a malicious intermediate router may attempt to insert additional ASes into the path. For example,
even though the advertised route goes directly from E to P ,
E may divert the traffic to another malicious AS, who then
forwards it to P . As long as the packets eventually pass
through P , the basic protocol of Section 3.2 will not detect
the deviation. However, in the extended protocol described
in Section 3.3, V specifies which incoming interface will be
used and thus where the secrets will be stored. Packets then
must arrive to P through the specified predecessor. Secrets
arriving to P through the wrong incoming interface will not
generate the expected responses.
The only exception is the case when the route contains
multiple consecutive malicious ASes who are willing to lie
to the verifier. They can insert additional ASes between
themselves without being detected. Moreover, a malicious
router can always copy all passing packets and share them
with the adversary. It is not clear whether such eavesdropping attacks can be detected using any verification method
that does not give the verifier complete access to the internal state of all routers on the data path, including those
controlled by malicious entities.
5.
EVALUATION
Experimental setup. We built a prototype implementation of the prover and verifier within the Click 1.5.0 software
router [13] on top of a modified Linux 2.6.16.13 kernel. Our
verifier and prover run on Pentium III/933 machines with
896 MB of RAM with two 64-bit AceNIC Gigabit Ethernet
cards in 64-bit 33 MHz PCI slots. Each end host had one
AceNIC Ethernet card and 512 MB of RAM but were otherwise the same. The machines are all connected as shown in
Fig. 3 via crossover cables. All experiments were run using
both Fast (100 Mbps) and Gigabit (1 Gbps) Ethernet links.
Verification was performed only on the traffic heading from
the client to the server. We vary the probe rate, or percentage of secret-tagged packets, and measure throughput,
latency, and CPU utilization.
Implementation. In our protocol, we chose to use 8-byte
tags and 4-byte return IP addresses, which are broken up
into six 16-bit pieces. We insert each piece into the 16-bit
fragmentation field of the header of each non-fragmented IP
packet. Since packets are rarely fragmented [21], this field
should almost always be available for tagging.
Our prover maintains a two-level sparse table of the most
recent tag fragments it has received from a particular verifier
indexed by the route identifier, which it uses to reconstruct
the tag and the return address, which is then looked up in
the secret table. This requires 12-byte route-specific state at
the prover to temporarily store fragments until the tag and
the address are reconstructed. This short-lived state is an
implementation artifact and not a necessary part of the protocol. We discuss alternative implementations in Section 6.
Since generating responses to shared secrets is the main
bottleneck at the prover, splitting each secret over six data
packets reduces the prover’s workload by decreasing the average number of responses per packet on the route that is
being verified. It also makes a given set of shared secrets last
six times longer before a new secret set has to be generated.
Because tags are broken up into multiple fragments which
are inserted into consecutive packets, the loss or misrouting
of any secret fragment is equivalent to losing or misrouting
the entire secret. Moreover, this implementation is sensitive
to packet reordering. If packets tagged with secret fragments
are reordered, the reassembled secret will likely be invalid
and thus generate no reply. Although reordering is rather
infrequent [18], our statistical analysis must still take it into
account. Assuming that any packet drop or reorder causes
a secret to be invalid, we redefine θn from (1) to
(1 − )(1 − δ)6 (1 − δ)(1 − φ)6 n ≡ θn n
where φ is the probability of packet reordering between the
verifier and prover and (1 − δ)6 represents the probability
that all tag fragments reach the prover. In Section 6, we discuss an alternative that is not sensitive to packet reordering.
The prover and verifier consist of a kernel-level Click component to handle packet modification and preliminary checks,
and a user-level process to transmit secrets to and from the
kernel-level component, track tag usage, and reply to secret tags. The prover and verifier start off by sharing a set
of 312,500 secret tuples. The prover then loads the entire
secret set, while the verifier buffers approximately 20,000
secret tuples. As the verifier tags packets and uses up secret tuples during route verification, the user-level process
pushes more to the verifier’s buffer via the /proc filesystem
interface. The verifier uses RC4-based frandom [2] after discarding the first 256 bytes to randomly decide which packets
to tag with secrets and to generate random secrets.
When a packet arrives at the prover’s router, it is checked
to see if the fragmentation offset is zero and the “more fragments” flag is disabled. A tag fragment is extracted from
the fragmentation identifier field and inserted into the appropriate sliding window buffer, which is then looked up in
a 8 MB 4-way cuckoo hash table [7] using four hash functions [11]. A cuckoo hash table guarantees constant-time
lookups at the cost of linear space overhead. Since such
a table could potentially be extremely large, we allocate it
within a user-level process. However, every tag then must
be copied from kernel memory to the user-level process’s
memory in order to be checked against the table of secrets.
Our implementation avoids many of these copies through a
small 1 MB Bloom filter [3] with four hash function within
our kernel-level prover. Through an extra set of lookups, it
can filter out many random tags, thus significantly reducing
our memory-copying overhead. Tags that pass the smaller
filter are copied via /proc to the user-level process, which
then checks the hash table. For every tag that the hash
table classifies as a secret, the user-level prover generates a
UDP packet with the appropriate response as a payload and
sends it to the listed IP address.
The verifier listens for incoming UDP responses and checks
them against the expected response set, which we implement using a counting Bloom filter [6], a generalization of
2.0.0.2
Client
Pentium III-933
512 MB RAM
2.0.0.1
1.0.0.2
Verifier
Pentium III-933
896 MB RAM
1.0.0.3
3.0.0.1
Prover
Pentium III-933
896 MB RAM
3.0.0.3
Server
Pentium III-933
512 MB RAM
Figure 3: Our testbed.
Figure 4: Average throughput of running an unmodified Click router compared to our protocol with a
varying proportion of probe packets.
Figure 5: Average round-trip time of a packet.
the Bloom filter that allows probabilistic element removal.
Alternatively, we could have used a larger, non-probabilistic
hash set to store all secrets.
Results. To estimate the overhead of our protocol, we used
Iperf 2.0.2 [10] to measure the average throughput of twentyone 30-second unidirectional TCP bulk data transfers over
both Gigabit (1 Gbps) and Fast Ethernet (100 Mbps) links.
Since Iperf was continuously transferring data as quickly as
possible, our packets were almost all 1500 bytes in length.
Fig. 4 shows the average throughput achieved with different proportions of probe packets. Using the unmodified
Click router, we were able to forward approximately 624.1
and 94.7 Mbps over Gigabit and Fast Ethernet links, respectively. Performing verification on Fast Ethernet resulted in
Figure 6: Average CPU utilization of our verifier
and prover.
no overhead regardless of the proportion of probe packets.
This is likely due to our machines being sufficiently powerful
to handle verification at line speed.
We then repeated our experiments on Gigabit Ethernet
links. Verification with a 1%, 2% and 5% probe ratio resulted in, respectively, throughputs of 621.7, 608.8, and 586
Mbps, overheads of 0.38%, 2.45%, and 6.10% over the standalone Click router. In our experiments, every packet is
tagged and thus must be inspected by the prover, but increasing the probe rate increases the number of UDP responses that the prover must send back. Disabling the
prover and running just the verifier with a 5% probe ratio, we achieved an average throughput equivalent to that
of running two unmodified Click routers. This indicates that
the overhead of random tagging is very small even when all
packets are being tagged, and that the overhead is mainly
caused by responding to probe packets on the prover side.
Another significant contributor to the overhead is the transfer of tags from kernel to user memory, which we solved
with a smaller, kernel-level Bloom filter. In the throughput experiments, we used a 1 MB filter to hold 312,500 secrets. Decreasing the size of the filter to 256 KB dropped
the throughput of 1% probing by about 13.7%.
We used 3000 ping requests to measure the average roundtrip latency of the packets traveling through routers running
our route verification protocol. On 100 Mbps and 1 Gbps
links, our protocol had no statistically significant impact on
latency, even when we increased the probe rate.
Finally, we measured the system load of running our protocol. We used the top utility to check CPU utilization
of the Click kernel thread and our user-level verifier and
prover every 3 seconds for 180 seconds while running our
throughput experiments. Fig. 6 shows the CPU overhead
of our protocol. Running the unmodified Click router uses
about 70 to 72% of the CPU. Running our modified version
of the Click router results in a 70 to 74% CPU utilization
on both the prover and verifier when the probe rate is decreased from 5% to 1%, respectively. Although it may seem
odd for the CPU utilization of our protocol to drop as the
probe rate increases, note that the Click router itself is more
CPU-intensive than our protocol; probing 1% of the packets incurs an overhead of at most 6.5% over the standalone
Click router even when the throughput is roughly the same.
We believe that the throughput of the Click router directly
corresponds to its CPU utilization as well; an unmodified
Click router on a 100 Mbps connection used up about 23 to
25% of a CPU’s time. Since increasing the probe rate causes
throughput to decline, which causes a decrease in the Click
router’s workload, we believe that this results in an overall
decline in CPU utilization as the probe rate increases.
6.
ALTERNATIVE APPROACHES
Set of secrets. In our prototype implementation, we chose
to have the prover store its secrets within a 4-way cuckoo
hash table [7] in order to guarantee constant-time lookups (k
i
hash functions) at the expense of O(|SR
P |) space complexity.
If space rather than computational power is the prover’s
bottleneck, the prover could distribute secrets of the form
(si , Hk (si ), Hk (Hk (si ))), where Hk is a keyed cryptographic
hash function and k is a secret key known only to the prover.
The prover then only has to remember k for all verifications
within a generation. When a verifier tags a packet, it either
embeds two random numbers or (si , Hk (si )). The prover
identifies a pair as a shared secret by applying Hk to the
first element and checking whether this equals the second
element. Since only the prover knows k, no one else can
distinguish a secret pair from a pair of random numbers.
The prover responds by sending back Hk (Hk (si )) in a similar
fashion as in the original protocol.
While conceptually attractive, this approach introduces
cryptographic operations into the critical path of the prover’s
border routers. In our informal tests, we found that computing a SHA-1 hash over an 8-byte buffer on one of our
Pentium III/933 router machines was approximately 60-70
times slower than the hash functions we used for our secret
table [11]. Moreover, since verifiers cannot generate secrets
on their own, entire sets of secrets have to be distributed
during the setup phase and stored on the verifiers’ routers.
Tagging packets. In our implementation, we embed 16-bit
tag fragments in consecutive packets. This method is sensitive to packet re-ordering and forces the prover to maintain
temporary route-specific state to rebuild the original tag.
An alternative is to use IP options to embed whole tags.
This avoids reordering issues and removes any route-specific
state from the prover. In addition, since IP options are
variable-sized, multiple verifiers along a given route could
simultaneously perform verification using the same packet.
Adding IP options, however, increases the size of packets
and may cause larger packets to become fragmented. More
importantly, packets with IP options will be routed via the
slow path on all routers, significantly impacting throughput.
paths of data traffic. The main features of our protocol are
its scalability and robustness against malicious routers on
the path. At the cost of an infrequent offline setup phase,
our protocol enables an autonomous system to prove both its
and its predecessor’s presence on the data path while handling high traffic loads and without maintaining a secure
channel with each verifier. Our evaluation of the prototype
implementation demonstrates that the impact of our protocol on router performance is small.
Acknowledgements. This material is based upon work
supported by the National Science Foundation under grants
CNS-0509033, CNS-0509338, IIS-0534198 and CyberTrust
award 0430510, and the Department of Defense under Contract No. H98230-05-C-1650.
8.[1] I.REFERENCES
Avramopoulos and J. Rexford. Stealth probing: Efficient
[2]
[3]
[4]
[5]
[6]
[7]
[8]
[9]
[10]
[11]
[12]
[13]
[14]
[15]
[16]
[17]
[18]
[19]
[20]
[21]
[22]
7.
CONCLUSIONS
We have presented a lightweight protocol for verifying consistency between Internet route advertisements and actual
[23]
data-plane security for IP routing. In Proc. USENIX Annual
Technical Conference, 2006.
E. Billauer. Frandom. http://frandom.sourceforge.net, 2007.
B. H. Bloom. Space/time trade-offs in hash coding with
allowable errors. CACM, 13(7):422–426, 1970.
Cisco Systems. Cisco CRS-1 carrier routing system.
http://www.cisco.com/application/pdf/en/us/guest/products/
ps5763/c1031/cdccont 0900aecd800f8118.pdf.
M. H. DeGroot. Probability and Statistics. Addison-Wesley,
1986.
L. Fan, P. Cao, J. Almeida, and A. Z. Broder. Summary cache:
a scalable wide-area web cache sharing protocol. IEEE/ACM
Trans. Netw., 8(3):281–293, 2000.
D. Fotakis, R. Pagh, P. Sanders, and P. G. Spirakis. Space
efficient hash tables with worst case constant access time. In
Proc. STACS, 2003.
S. Goldberg, D. Xiao, B. Barak, and J. Rexford. Measuring
path quality in the presence of adversaries.
http://www.princeton.edu/∼goldbe/FDFL-sc.pdf, 2007.
G. Goodell, W. Aiello, T. Griffin, J. Ioannidis, P. McDaniel,
and A. Rubin. Working around BGP: An incremental approach
to improving security and accuracy of interdomain routing. In
Proc. NDSS, 2003.
Iperf. The TCP/UDP bandwidth measurement tool.
http://dast.nlanr.net/Projects/Iperf/, 2005.
B. Jenkins. Hash functions and block ciphers.
http://www.burtleburtle.net/bob/hash, 2006.
S. Kent, C. Lynn, and K. Seo. Secure Border Gateway protocol
(Secure-BGP). IEEE Journal on Selected Areas in
Communications, 18(4), 2000.
E. Kohler, R. Morris, B. Chen, J. Jannotti, and M. F.
Kaashoek. The Click modular router. ACM Trans. Comput.
Syst., 18(3):263–297, 2000.
R. Mahajan, D. Wetherall, and T. Anderson. Understanding
BGP misconfigurations. In Proc. SIGCOMM, 2002.
Z. Mao, J. Rexford, J. Wang, and R. Katz. Towards an
accurate AS-level traceroute tool. In Proc. SIGCOMM, 2003.
A. Mizrak, Y. Cheng, K. Marzullo, and S. Savage. Fatih:
Detecting and isolating malicious routers. In Proc. DSN, 2005.
V. Padmanabhan and D. Simon. Secure traceroute to detect
faulty or malicious routing. SIGCOMM Comp. Comm.
Review, 33(1):77–82, 2003.
V. Paxson. End-to-end internet packet dynamics. IEEE/ACM
Trans. Netw., 7(3):277–292, 1999.
S. Savage, D. Wetherall, A. Karlin, and T. Anderson. Practical
network support for IP traceback. SIGCOMM Comp. Comm.
Rev., 30(4):295–306, 2000.
A. C. Snoeren, C. Partridge, L. A. Sanchez, C. E. Jones,
F. Tchakountio, B. Schwartz, S. T. Kent, and W. T. Strayer.
Single-packet IP traceback. IEEE/ACM Trans. Netw.,
10(6):721–734, 2002.
I. Stoica and H. Zhang. Providing guaranteed services without
per-flow management. In Proc. SIGCOMM, 1999.
L. Subramanian, V. Roth, I. Stoica, S. Shenker, and R. Katz.
Listen and Whisper: Security mechanisms for BGP. In Proc.
NSDI, 2004.
T. Wan, E. Kranakis, and P. van Oorschot. Pretty secure BGP
(psBGP). In Proc. NDSS, 2005.
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertising