IEEE BE-BTECH NS2 PROJECT@ DREAMWEB TECHNO SOLUTION

Privacy-Preserving and Truthful Detection
of Packet Dropping Attacks in Wireless
Ad Hoc Networks
Tao Shu and Marwan Krunz, Fellow, IEEE
Abstract—Link error and malicious packet dropping are two sources for packet losses in multi-hop wireless ad hoc network. In this
paper, while observing a sequence of packet losses in the network, we are interested in determining whether the losses are caused by
link errors only, or by the combined effect of link errors and malicious drop. We are especially interested in the insider-attack case,
whereby malicious nodes that are part of the route exploit their knowledge of the communication context to selectively drop a small
amount of packets critical to the network performance. Because the packet dropping rate in this case is comparable to the channel error
rate, conventional algorithms that are based on detecting the packet loss rate cannot achieve satisfactory detection accuracy. To
improve the detection accuracy, we propose to exploit the correlations between lost packets. Furthermore, to ensure truthful calculation
of these correlations, we develop a homomorphic linear authenticator (HLA) based public auditing architecture that allows the detector
to verify the truthfulness of the packet loss information reported by nodes. This construction is privacy preserving, collusion proof, and
incurs low communication and storage overheads. To reduce the computation overhead of the baseline scheme, a packet-block-based
mechanism is also proposed, which allows one to trade detection accuracy for lower computation complexity. Through extensive
simulations, we verify that the proposed mechanisms achieve significantly better detection accuracy than conventional methods such
as a maximum-likelihood based detection.
Index Terms—Packet dropping, secure routing, attack detection, homomorphic linear signature, auditing
Ç
1 INTRODUCTION
IN a multi-hop wireless network, nodes cooperate in relay-
ing/routing traffic. An adversary can exploit this coopera-
tive nature to launch attacks. For example, the adversary
may first pretend to be a cooperative node in the route dis-
covery process. Once being included in a route, the adver-
sary starts dropping packets. In the most severe form, the
malicious node simply stops forwarding every packet
received from upstream nodes, completely disrupting the
path between the source and the destination. Eventually,
such a severe denial-of-service (DoS) attack can paralyze the
network by partitioning its topology.
Even though persistent packet dropping can effectively
degrade the performance of the network, from the attacker’s
standpoint such an “always-on” attack has its disadvan-
tages. First, the continuous presence of extremely high
packet loss rate at the malicious nodes makes this type of
attack easy to be detected [25]. Second, once being detected,
these attacks are easy to mitigate. For example, in case the
attack is detected but the malicious nodes are not identified,
one can use the randomized multi-path routing algorithms
[28], [29] to circumvent the black holes generated by the
attack, probabilistically eliminating the attacker’s threat. If
the malicious nodes are also identified, their threats can be
completely eliminated by simply deleting these nodes from
the network’s routing table.
A malicious node that is part of the route can exploit its
knowledge of the network protocol and the communication
context to launch an insider attack—an attack that is intermit-
tent, but can achieve the same performance degradation
effect as a persistent attack at a much lower risk of being
detected. Specifically, the malicious node may evaluate the
importance of various packets, and then drop the small
amount that are deemed highly critical to the operation of
the network. For example, in a frequency-hopping network,
these could be the packets that convey frequency hopping
sequences for network-wide frequency-hopping synchroni-
zation; in an ad hoc cognitive radio network, they could be
the packets that carry the idle channel lists (i.e., white
spaces) that are used to establish a network-wide control
channel. By targeting these highly critical packets, the
authors in [21], [24], [25] have shown that an intermittent
insider attacker can cause significant damage to the network
with low probability of being caught. In this paper, we are
interested in combating such an insider attack. In particular,
we are interested in the problem of detecting the occurrence
of selective packet drops and identifying the malicious
node(s) responsible for these drops.
Detecting selective packet-dropping attacks is extremely
challenging in a highly dynamic wireless environment. The
difficulty comes from the requirement that we need to not
only detect the place (or hop) where the packet is dropped,
but also identify whether the drop is intentional or
unintentional. Specifically, due to the open nature of wireless
medium, a packet drop in the network could be caused by
harsh channel conditions (e.g., fading, noise, and interfer-
ence, a.k.a., link errors), or by the insider attacker. In an open
T. Shu is with the Department of Computer Science and Engineering,
Oakland University, Rochester, MI. E-mail: shu@oakland.edu.
M. Krunz is with the Department of Electrical and Computer Engineering,
the University of Arizona, Tucson, AZ. E-mail: krunz@email.arizona.edu.
Manuscript received 13 Apr. 2013; revised 18 Mar. 2014; accepted 3 June
2014. Date of publication 12 June 2014; date of current version 2 Mar. 2015.
For information on obtaining reprints of this article, please send e-mail to:
reprints@ieee.org, and reference the Digital Object Identifier below.
Digital Object Identifier no. 10.1109/TMC.2014.2330818
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 14, NO. 4, APRIL 2015 813
1536-1233 ß 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

wireless environment, link errors are quite significant, and
may not be significantly smaller than the packet dropping
rate of the insider attacker. So, the insider attacker can cam-
ouflage under the background of harsh channel conditions.
In this case, just by observing the packet loss rate is not
enough to accurately identify the exact cause of a packet loss.
The above problem has not been well addressed in the
literature. As discussed in Section 2, most of the related
works preclude the ambiguity of the environment by
assuming that malicious dropping is the only source of
packet loss, so that there is no need to account for the impact
of link errors. On the other hand, for the small number of
works that differentiate between link errors and malicious
packet drops, their detection algorithms usually require the
number of maliciously-dropped packets to be significantly
higher than link errors, in order to achieve an acceptable
detection accuracy.
In this paper, we develop an accurate algorithm for
detecting selective packet drops made by insider attackers.
Our algorithm also provides a truthful and publicly verifi-
able decision statistics as a proof to support the detection
decision. The high detection accuracy is achieved by exploit-
ing the correlations between the positions of lost packets, as
calculated from the auto-correlation function (ACF) of the
packet-loss bitmap—a bitmap describing the lost/received
status of each packet in a sequence of consecutive packet
transmissions. The basic idea behind this method is that
even though malicious dropping may result in a packet loss
rate that is comparable to normal channel losses, the stochas-
tic processes that characterize the two phenomena exhibit
different correlation structures (equivalently, different pat-
terns of packet losses). Therefore, by detecting the correla-
tions between lost packets, one can decide whether the
packet loss is purely due to regular link errors, or is a com-
bined effect of link error and malicious drop. Our algorithm
takes into account the cross-statistics between lost packets to
make a more informative decision, and thus is in sharp con-
trast to the conventional methods that rely only on the distri-
bution of the number of lost packets.
The main challenge in our mechanism lies in how to guar-
antee that the packet-loss bitmaps reported by individual
nodes along the route are truthful, i.e., reflect the actual sta-
tus of each packet transmission. Such truthfulness is essen-
tial for correct calculation of the correlation between lost
packets. This challenge is not trivial, because it is natural for
an attacker to report false information to the detection algo-
rithm to avoid being detected. For example, the malicious
node may understate its packet-loss bitmap, i.e., some pack-
ets may have been dropped by the node but the node reports
that these packets have been forwarded. Therefore, some
auditing mechanism is needed to verify the truthfulness of
the reported information. Considering that a typical wireless
device is resource-constrained, we also require that a user
should be able to delegate the burden of auditing and detec-
tion to some public server to save its own resources.
Our solution to the above public-auditing problem is con-
structed based on the homomorphic linear authenticator
(HLA) cryptographic primitive [2], [3], [27], which is basi-
cally a signature scheme widely used in cloud computing
and storage server systems to provide a proof of storage
from the server to entrusting clients [30]. However, direct
application of HLA does not solve our problem well, mainly
because in our problem setup, there can be more than one
malicious node along the route. These nodes may collude
(by exchanging information) during the attack and when
being asked to submit their reports. For example, a packet
and its associated HLA signature may be dropped at an
upstream malicious node, so a downstream malicious node
does not receive this packet and the HLA signature from the
route. However, this downstream attacker can still open a
back-channel to request this information from the upstream
malicious node. When being audited, the downstream mali-
cious node can still provide valid proof for the reception of
the packet. So packet dropping at the upstream malicious
node is not detected. Such collusion is unique to our prob-
lem, because in the cloud computing/storage server sce-
nario, a file is uniquely stored at a single server, so there are
no other parties for the server to collude with. We show that
our new HLA construction is collusion-proof.
Our construction also provides the following new fea-
tures. First, privacy-preserving: the public auditor should
not be able to decern the content of a packet delivered on
the route through the auditing information submitted by
individual hops, no matter how many independent
reports of the auditing information are submitted to the
auditor. Second, our construction incurs low communica-
tion and storage overheads at intermediate nodes. This
makes our mechanism applicable to a wide range of wire-
less devices, including low-cost wireless sensors that
have very limited bandwidth and memory capacities.
This is also in sharp contrast to the typical storage-server
scenario, where bandwidth/storage is not considered an
issue. Last, to significantly reduce the computation over-
head of the baseline constructions so that they can be
used in computation-constrained mobile devices, a
packet-block-based algorithm is proposed to achieves
scalable signature generation and detection. This mecha-
nism allows one to trade detection accuracy for lower
computation complexity.
The remainder of this paper is organized as follows. In
Section 2 we review the related work. The system/adver-
sary models and problem statement are described in Section
3. We present the proposed scheme and analyze its security
performance and overheads in Section 4. The low-computa-
tion-overhead block-based algorithm is proposed in Section
5. Simulation results are presented in Section 6, and we con-
clude the paper in Section 7.
2 RELATED WORK
Depending on how much weight a detection algorithm gives
to link errors relative to malicious packet drops, the related
work can be classified into the following two categories.
The first category aims at high malicious dropping rates,
where most (or all) lost packets are caused by malicious
dropping. In this case, the impact of link errors is ignored.
Most related work falls into this category. Based on the
methodology used to identify the attacking nodes, these
works can be further classified into four sub-categories. The
first sub-category is based on credit systems [9], [34], [10]. A
credit system provides an incentive for cooperation. A node
receives credit by relaying packets for others, and uses its
814 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 14, NO. 4, APRIL 2015

credit to send its own packets. As a result, a maliciously
node that continuous to drop packets will eventually
deplete its credit, and will not be able to send its own traffic.
The second sub-category is based on reputation systems
[12], [8], [14], [19], [20], [11], [4]. A reputation system relies
on neighbors to monitor and identify misbehaving nodes. A
node with a high packet dropping rate is given a bad repu-
tation by its neighbors. This reputation information is prop-
agated periodically throughout the network and is used as
an important metric in selecting routes. Consequently, a
malicious node will be excluded from any route. The third
sub-category of works relies on end-to-end or hop-to-hop
acknowledgements to directly locate the hops where pack-
ets are lost [18], [22], [23], [5], [6], [32]. A hop of high packet
loss rate will be excluded from the route. The fourth sub-
category addresses the problem using cryptographic meth-
ods. For example, the work in [17] utilizes Bloom filters to
construct proofs for the forwarding of packets at each node.
By examining the relayed packets at successive hops along
a route, one can identify suspicious hops that exhibit high
packet loss rates. Similarly, the method in [16], [33] traces
the forwarding records of a particular packet at each inter-
mediate node by formulating the tracing problem as a
Renyi-Ulam game. The first hop where the packet is no lon-
ger forwarded is considered a suspect for misbehaving.
The second category targets the scenario where the num-
ber of maliciously dropped packets is significantly higher
than that caused by link errors, but the impact of link errors
is non-negligible. Certain knowledge of the wireless channel
is necessary in this case. The authors in [26] proposed to
shape the traffic at the MAC layer of the source node
according to a certain statistical distribution, so that inter-
mediate nodes are able to estimate the rate of received traf-
fic by sampling the packet arrival times. By comparing the
source traffic rate with the estimated received rate, the
detection algorithm decides whether the discrepancy in
rates, if any, is within a reasonable range such that the dif-
ference can be considered as being caused by normal chan-
nel impairments only, or caused by malicious dropping,
otherwise. The works in [13] and [31] proposed to detect
malicious packet dropping by counting the number of lost
packets. If the number of lost packets is significantly larger
than the expected packet loss rate made by link errors, then
with high probability a malicious node is contributing to
packet losses.
All methods mentioned above do not perform well when
malicious packet dropping is highly selective. More specifi-
cally, for the credit-system-based method, a malicious node
may still receive enough credits by forwarding most of the
packets it receives from upstream nodes. Similarly, in the
reputation-based approach, the malicious node can main-
tain a reasonably good reputation by forwarding most of
the packets to the next hop. While the Bloom-filter scheme
is able to provide a packet forwarding proof, the correctness
of the proof is probabilistic and it may contain errors. For
highly selectively attacks (low packet-dropping rate), the
intrinsic error rate of Bloom filer significantly undermines
its detection accuracy. As for the acknowledgement-based
method and all the mechanisms in the second category,
merely counting the number of lost packets does not give a
sufficient ground to detect the real culprit that is causing
packet losses. This is because the difference in the number
of lost packets between the link-error-only case and the
link-error-plus-malicious-dropping case is small when the
attacker drops only a few packets. Consequently, the detec-
tion accuracy of these algorithms deteriorates when mali-
cious drops become highly selective.
Our study targets the challenging situation where link
errors and malicious dropping lead to comparable packet
loss rates. The effort in the literature on this problem has
been quite preliminary, and there is a few related works.
Note that the cryptographic methods proposed in [24] to
counter selective packet jamming target a different issue
than the detection problem studied in this paper. The meth-
ods in [24] delay a jammer from recognizing the significance
of a packet after the packet has been successfully transmit-
ted, so that there is no time for the jammer to conduct jam-
ming based on the content/importance of the packet.
Instead of trying to detect any malicious behavior, the
approach in [24] is proactive, and hence incurs overheads
regardless of the presence or absence of attackers.
3 SYSTEM MODELS AND PROBLEM STATEMENT
3.1 Network and Channel Models
Consider an arbitrary path PSD in a multi-hop wireless ad
hoc network, as shown in Fig. 1. The source node S continu-
ously sends packets to the destination node D through inter-
mediate nodes n1; . . . ; nK, where ni is the upstream node of
niþ1, for 1 i K À 1. We assume that S is aware of the
route PSD, as in Dynamic Source Routing (DSR) [15]. If DSR
is not used, S can identify the nodes in PSD by performing a
traceroute operation. Here we mainly focus on static or
quasi-static wireless ad hoc networks, i.e., we assume that
the network topology and link characteristics remain
unchanged for a relatively long period of time. Example net-
works include wireless mesh networks (WMNs) and ad hoc
networks formed in nomadic computing. Extension to a
highly mobile environment is out of our scope and will be
considered in the future work.
We model the wireless channel of each hop along PSD as
a random process that alternates between good and bad
states. Packets transmitted during the good state are suc-
cessful, and packets transmitted during the bad state are
lost. In contrast to the classical Gilbert-Ellioit (GE) channel
model, here we do not assume any Markovian property on
the channel behavior. We only require that the sequence of
sojourn times for each state follows a stationary distribu-
tion, and the autocorrelation function of the channel state,
say fcðiÞ, where i is the time lag in packets, is also
Fig. 1. Network and attack model.
SHU AND KRUNZ: PRIVACY-PRESERVING AND TRUTHFUL DETECTION OF PACKET DROPPING ATTACKS IN WIRELESS AD HOC... 815

stationary. Here we limit our study to quasi-static networks,
whereby the path PSD remains unchanged for a relatively
long time, so that the link error statistics of the wireless
channel is a wide-sense stationary (WSS) random process
(i.e., fcðiÞ is stationary). Detecting malicious packet drops
may not be a concern for highly mobile networks, because
the fast-changing topology of such networks makes route
disruption the dominant cause for packet losses. In this
case, maintaining stable connectivity between nodes is a
greater concern than detecting malicious nodes. The func-
tion fcðiÞ can be calculated using the probing approach in
[1]. In brief, a sequence of M packets are transmitted consec-
utively over the channel. By observing whether the trans-
missions are successful or not, the receiver obtains a
realization of the channel state ða1; . . . ; aMÞ, where
aj 2 f0; 1g for j ¼ 1; . . . ; M. In this sequence, “1” denotes
the packet was successfully received, and “0” denotes the
packet was dropped. fcðiÞ is derived by computing the
autocorrelation function of this sample sequence:
fcðiÞ def
¼ Efajajþig for i ¼ 0; . . . ; M, where the expectation is
calculated over all transmitted packets j ¼ 1; . . . ; M. This
autocorrelation function describes the correlation between
packet transmissions (successful/lost) at different times, as
a function of the time lag. The time invariant nature of fc is
guaranteed by the WSS assumption of the wireless channel.
The measurement of fcðiÞ can take place online or offline. A
detailed discussion on how fcðiÞ is derived is out of the
scope of this paper, and we simply assume that this infor-
mation is given as input to our detection algorithm.
There is an independent auditor Ad in the network. Ad is
independent in the sense that it is not associated with any
node in PSD and does not have any knowledge of the secrets
(e.g., cryptographic keys) held by various nodes. The audi-
tor is responsible for detecting malicious nodes on demand.
Specifically, we assume S receives feedback from D when D
suspects that the route is under attack. Such a suspicion
may be triggered by observing any abnormal events, e.g., a
significant performance drop, the loss of multiple packets of
a certain type, etc. We assume that the integrity and authen-
ticity of the feedback from D to S can be verified by S using
resource-efficient cryptographic methods such as the Ellip-
tic Curve Digital Signature Algorithm (ECDSA). Once being
notified of possible attacks, S submits an attack-detection
request (ADR) to Ad. To facilitate its investigation, Ad needs
to collect certain information (elaborated on in the next sec-
tion) from the nodes on route PSD. We assume that each
such node must reply to Ad’s inquiry, otherwise the node
will be considered as misbehaving. We assume that normal
nodes will reply with truthful information, but malicious
nodes may cheat. At the same time, for privacy reasons, we
require that Ad cannot determine the content of the normal
packets delivered over PSD from the information collected
during the auditing.
3.2 Adversarial Model
The goal of the adversary is to degrade the network’s perfor-
mance by maliciously dropping packets while remaining
undetected. We assume that the malicious node has knowl-
edge of the wireless channel, and is aware of the algorithm
used for misbehavior detection. It has the freedom to choose
what packets to drop. For example, in the random-drop
mode, the malicious node may drop any packet with a small
probability pd. In the selective-mode, the malicious node only
drops packets of certain types. A combination of the two
modes may be used. We assume that any node on PSD can be
a malicious node, except the source and the destination. In
particular, there can be multiple malicious nodes on PSD.
We consider the following form of collusion between
malicious nodes: A covert communication channel may
exist between any two malicious nodes, in addition to the
path connecting them on PSD. As a result, malicious nodes
can exchange any information without being detected by Ad
or any other nodes in PSD. Malicious nodes can take advan-
tage of this covert channel to hide their misbehavior and
reduce the chance of being detected. For example, an
upstream malicious node may drop a packet on PSD, but
may secretely send this packet to a downstream malicious
node via the covert channel. When being investigated, the
downstream malicious node can provide a proof of the suc-
cessful reception of the packet. This makes the auditor
believe that the packet was successfully forwarded to the
downstream nodes, and not know that the packet was actu-
ally dropped by an upstream attacker.
3.3 Problem Statement
Under the system and adversary models defined above, we
address the problem of identifying the nodes on PSD that
drop packets maliciously. We require the detection to be
performed by a public auditor that does not have knowl-
edge of the secrets held by the nodes on PSD. When a mali-
cious node is identified, the auditor should be able to
construct a publicly verifiable proof of the misbehavior of
that node. The construction of such a proof should be pri-
vacy preserving, i.e., it does not reveal the original informa-
tion that is transmitted on PSD. In addition, the detection
mechanism should incur low communication and storage
overheads, so that it can be applied to a wide variety of
wireless networks.
4 PROPOSED DETECTION SCHEME
4.1 Overview
The proposed mechanism is based on detecting the correla-
tions between the lost packets over each hop of the path.
The basic idea is to model the packet loss process of a hop
as a random process alternating between 0 (loss) and 1 (no
loss). Specifically, consider that a sequence of M packets
that are transmitted consecutively over a wireless channel.
By observing whether the transmissions are successful or
not, the receiver of the hop obtains a bitmap ða1; . . . ; aMÞ,
where aj 2 f0; 1g for packets j ¼ 1; . . . ; M. The correlation
of the lost packet is calculated as the auto-correlation
function of this bitmap. Under different packet dropping
conditions, i.e., link-error versus malicious dropping, the
instantiations of the packet-loss random process should
present distinct dropping patterns (represented by the cor-
relation of the instance). This is true even when the packet
loss rate is similar in each instantiation. To verify this prop-
erty, in Fig. 2 we have simulated the auto-correlation func-
tions of two packet loss processes, one caused by 10 percent
link errors, and the other by 10 percent link errors plus
10 percent malicious uniformly-random packet dropping. It

can be observed that significant gap exists between these
two auto-correlation functions. Therefore, by comparing the
auto-correlation function of the observed packet loss pro-
cess with that of a normal wireless channel (i.e., fcðiÞ), one
can accurately identify the cause of the packet drops.
The benefit of exploiting the correlation of lost packets
can be better illustrated by examining the insufficiency of
the conventional method that relies only on the distribution
of the number of lost packets. More specifically, under the
conventional method, malicious-node detection is modeled
as a binary hypothesis test, where H0 is the hypothesis that
there is no malicious node in a given link (all packet losses
are due to link errors) and H1 denotes there is a malicious
node in the given link (packet losses are due to both link
errors and malicious drops). Let z be the observed number
of lost packets on the link during some interval t. Then,
z ¼
x; under H0 ðno malicious nodesÞ
x þ y; under H1ðthere is a malicious nodeÞ;

(1)
where x and y are the numbers of lost packets caused by
link errors and by malicious drops, respectively. Both x
and y are random variables. Let the probability density
functions of z conditioned on H0 and on H1 be h0ðzÞ and
h1ðzÞ, respectively, as shown in Fig. 3a. We are interested
in the maximum-uncertainty scenario where the a priori
probabilities are given by PrfH0g ¼ PrfH1g ¼ 0:5, i.e., the
auditor has no prior knowledge of the distributions of H0
and H1 to make any biased decision regarding the pres-
ence of malicious nodes. Let the false-alarm and miss-
detection probabilities be Pfa and Pmd, respectively. The
optimal decision strategy that minimizes the total detection
error Pdedef¼0:5ðPfa þ PmdÞ is the maximum-likelihood
(ML) algorithm:
if z zth; accept H0;
otherwise; accept H1;

(2)
where the threshold zth is the solution to the equation
h0ðzthÞ ¼ h1ðzthÞ. Under this strategy, Pfa and Pmd are the
areas of the shaded regions shown in Fig. 3a, respectively.
The problem with this mechanism is that, when the mean of
y is small, h1ðzÞ and h0ðzÞ are not sufficiently separated, lead-
ing to large Pfa and Pmd, as shown in Fig. 3b. This observation
implies that when malicious packet drops are highly selec-
tive, counting the number of lost packets is not sufficient to
accurately differentiate between malicious drops and link
errors. For such a case, we use the correlation between lost
packets to form a more informative decision statistic.
To correctly calculate the correlation between lost pack-
ets, it is critical to enforce a truthful packet-loss bitmap
report by each node. We use HLA cryptographic primitive
for this purpose. The basic idea of our method is as follows.
An HLA scheme allows the source, which has knowledge of
the HLA secret key, to generate HLA signatures s1; . . . ; sM
for M independent messages r1; . . . ; rM, respectively. The
source sends out the ri’s and si’s along the route. The HLA
signatures are made in such a way that they can be used as
the basis to construct a valid HLA signature for any arbi-
trary linear combination of the messages,
PM
i¼1 ciri, without
the use of the HLA secret key, where ci’s are randomly cho-
sen coefficients. A valid HLA signature for
PM
i¼1 ciri can be
constructed by a node that does not have knowledge of the
secret HLA key if and only if the node has full knowledge
of s1; . . . ; sM. So, if a node with no knowledge of the HLA
secret key provides a valid signature for
PM
i¼1 ciri, it implies
that this node must have received all the signatures
s1; . . . ; sM. Our construction ensures that si and ri are sent
together along the route, so that knowledge of s1; . . . ; sM
also proves that the node must have received r1; . . . ; rM.
Our detection architecture consists of four phases: setup,
packet transmission, audit, and detection. We elaborate on
these phases in the next section.
4.2 Scheme Details
4.2.1 Setup Phase
This phase takes place right after route PSD is established,
but before any data packets are transmitted over the route.
In this phase, S decides on a symmetric-key crypto-system
ðencryptkey; decryptkeyÞ and K symmetric keys key1; . . . ;
keyK, where encryptkey and decryptkey are the keyed encryp-
tion and decryption functions, respectively. S securely dis-
tributes decryptkey and a symmetric key keyj to node nj on
PSD, for j ¼ 1; . . . ; K. Key distribution may be based on the
public-key crypto-system such as RSA: S encrypts keyj
using the public key of node nj and sends the cipher text to
nj. nj decrypts the cipher text using its private key to obtain
keyj. S also announces two hash functions, H1 and HMAC
key , to
all nodes in PSD. H1 is unkeyed while HMAC
key is a keyed hash
function that will be used for message authentication pur-
poses later on.
Besides symmetric key distribution, S also needs to set
up its HLA keys. Let e : G Â G ! GT be a computable bilin-
ear map with multiplicative cyclic group G and support Zp,
where p is the prime order of G, i.e., for all a, b 2 G and q1,
Fig. 2. Comparison of correlation of lost packets.
Fig. 3. Insufficiency of conventional detection algorithms when malicious
packet drops are highly selective.

q2 2 Zp, eðaq1 ; bq2 Þ ¼ eða; bÞq1q2
. Let g be a generator of G.
H2ð:Þ is a secure map-to-point hash function: f0; 1gÃ
! G,
which maps strings uniformly to G. S chooses a random
number x 2 Zp and computes v ¼ gx
. Let u be another gen-
erator of G. The secret HLA key is sk ¼ x and the public
HLA key is a tuple pk ¼ ðv; g; uÞ.
4.2.2 Packet Transmission Phase
After completing the setup phase, S enters the packet trans-
mission phase. S transmits packets to PSD according to the
following steps.
Before sending out a packet Pi, where i is a sequence
number that uniquely identifies Pi, S computes
ri ¼ H1ðPiÞ and generates the HLA signatures of ri for
node nj, as follows:
sji ¼
Â
H2ðijjjÞuri
Ãx
; for j ¼ 1; . . . ; K; (3)
where jj denotes concatenation. These signatures are then
sent together with Pi to the route by using a one-way
chained encryption that prevents an upstream node from
deciphering the signatures intended for downstream nodes.
More specifically, after getting sji for j ¼ 1; . . . ; K, S itera-
tively computes the following:
~sKi ¼ encryptkeyK
ðsKiÞ;
tKi ¼ ~sKijjMACkeyK
ð~sKiÞ;
~sKÀ1i ¼ encryptkeyKÀ1
ðsKÀ1ijjtKiÞ;
tKÀ1i ¼ ~sKÀ1ijjMACkeyKÀ1
ð~sKÀ1iÞ;
..
.
~sji ¼ encryptkeyj ðsjijjtjþ1iÞ;
tji ¼ ~sjijjMACkeyj ð~sjiÞ;
..
.
~s1i ¼ encryptkey1
ðs1ijjt2iÞ;
t1i ¼ ~s1ijjMACkey1
ð~s1iÞ;
(4)
where the message authentication code (MAC) in each stage j
is computed according to the hash function HMAC
keyj
. After get-
ting t1i, S puts Pijjt1i into one packet and sends it to node n1.
When node n1 receives the packet from S, it extracts Pi,
~s1i, and MACkey1
ð~s1iÞ from the received packet. Then, n1
verifies the integrity of ~s1i by testing the following equality:
MACkey1
ð~s1iÞ ¼ HMAC
key1
ð~s1iÞ: (5)
If the test is true, then n1 decrypts ~s1i as follows:
decryptkey1
ð~s1iÞ ¼ s1ijjt2i: (6)
Then, n1 extracts s1i and t2i from the decrypted text. It stores
ri ¼ H1ðPiÞ and s1i in its proof-of-reception database for
future use. This database is maintained at every node on
PSD. It can be considered as a FIFO queue of size M, which
records the reception status for the most recent M packets
sent by S. Finally, n1 assembles Pijjt2i into one packet and
relays this packet to node n2. In case the test in (5) fails, n1
marks the loss of Pi in its proof-of-reception database and
does not relay the packet to n2.
The above process is repeated at every intermediate node
nj, j ¼ 1; . . . ; K. As a result, node nj obtains ri and its HLA
signature sji for every packet Pi that the node has received,
and it relays Pijjtjþ1i to the next hop on the route. The last
hop, i.e., node nK, only forwards Pi to the destination D. As
proved in Theorem 4 in Section 4.3, the special structure of
the one-way chained encryption construction in (4) dictates
that an upstream node on the route cannot get a copy of the
HLA signature intended for a downstream node, and thus
the construction is resilient to the collusion model defined
in Section 3.2. Note that here we consider the verification of
the integrity of Pi as an orthogonal problem to that of verify-
ing the tag tji. If the verification of Pi fails, node n1 should
also stop forwarding the packet and should mark it accord-
ingly in its proof-of-reception database.
4.2.3 Audit Phase
This phase is triggered when the public auditor Ad receives
an ADR message from S. The ADR message includes the id
of the nodes on PSD, ordered in the downstream direction,
i.e., n1; . . . ; nK, S’s HLA public key information pk ¼
ðv; g; uÞ, the sequence numbers of the most recent M packets
sent by S, and the sequence numbers of the subset of these
M packets that were received by D. Recall that we assume
the information sent by S and D is truthful, because detect-
ing attacks is in their interest. Ad conducts the auditing pro-
cess as follows.
Ad submits a random challenge vector ~cj ¼ ðcj1; . . . ; cjMÞ
to node nj, j ¼ 1; . . . ; K, where the elements cji’s are ran-
domly chosen from Zp. Without loss of generality, let the
sequence number of the packets recorded in the current
proof-of-reception database be P1; . . . ; PM , with PM being
the most recent packet sent by S. Based on the information
in this database, node nj generates a packet-reception bit-
map ~bj ¼ ðbj1; . . . ; bjMÞ, where bji ¼ 1 if Pi has been received
by nj, and bji ¼ 0 otherwise. Node nj then calculates the lin-
ear combination rðjÞ
¼
PM
i¼1;bji6¼0 cjiri and the HLA signature
for the combination as follows:
sðjÞ
¼
Y
i¼1;bji6¼0
s
cji
ji : (7)
Node nj submits ~bj, rðjÞ
, and sðjÞ
to Ad, as proof of the pack-
ets it has received.
Ad checks the validity of rðjÞ
and sðjÞ
by testing the follow-
ing equality:
eðsðjÞ
; gÞ ¼ e
YM
i¼1;bji6¼0
H2ðijjjÞcji
urðjÞ
; v
0
@
1
A: (8)
If the equality holds, then Ad accepts that node nj received
the packets as reflected in ~bj. Otherwise, Ad rejects ~bj and
judges that not all packets claimed in ~bj are actually received
by nj, so nj is a malicious node. We prove the correctness of
this auditing algorithm in Section 4.3.
Note that the above mechanism only guarantees that a
node cannot understate its packet loss, i.e., it cannot claim
the reception of a packet that it actually did not receive.
This mechanism cannot prevent a node from overly stating
its packet loss by claiming that it did not receive a packet
that it actually received. This latter case is prevented by
another mechanism discussed in the detection phase.

4.2.4 Detection Phase
The public auditor Ad enters the detection phase after receiv-
ing and auditing the reply to its challenge from all nodes on
PSD. The main tasks of Ad in this phase include the following:
detecting any overstatement of packet loss at each node, con-
structing a packet-loss bitmap for each hop, calculating the
autocorrelation function for the packet loss on each hop, and
deciding whether malicious behavior is present. More specif-
ically, Ad performs these tasks as follows.
Given the packet-reception bitmap at each node, ~b1; . . . ;
~bK, Ad first checks the consistency of the bitmaps for any
possible overstatement of packet losses. Clearly, if there is
no overstatement of packet loss, then the set of packets
received at node j þ 1 should be a subset of the packets
received at node j, for j ¼ 1; . . . ; K À 1. Because a normal
node always truthfully reports its packet reception, the
packet-reception bitmap of a malicious node that overstates
its packet loss must contradict with the bitmap of a normal
downstream node. Note that there is always at least one
normal downstream node, i.e., the destination D. So Ad only
needs to sequentially scan ~bj’s and the report from D to
identify nodes that are overstating their packet losses.
After checking for the consistency of ~bj’s, Ad starts con-
structing the per-hop packet-loss bitmap ~mj from ~bjÀ1 and
~bj. This is done sequentially, starting from the first hop from
S. In each step, only packets that are lost in the current hop
will be accounted for in mj. The packets that were not
received by the upstream node will be marked as “not lost”
for the underlying hop. Denoting the “lost” packet by 0 and
“not lost” by 1, ~mj can be easily constructed by conducting
a bit-wise complement-XOR operation of ~bjÀ1 and ~bj. For
example, consider the following simple case with three
intermediate nodes (four hops) on the route and with
M ¼ 10. Suppose that ~b1 ¼ ð0; 1; 1; 1; 1; 1; 1; 1; 0; 1Þ, ~b2 ¼
ð0; 1; 1; 1; 1; 1; 1; 1; 0; 1Þ, ~b3 ¼ ð0; 1; 0; 1; 1; 0; 1; 1; 0; 1Þ, and the
destination D reports that ~bD ¼ ð0; 1; 0; 1; 1; 0; 1; 1; 0; 1Þ. Then
the per-hop packet-loss bitmaps are given by ~m1 ¼
ð0; 1; 1; 1; 1; 1; 1; 1; 0; 1Þ, ~m2 ¼ ð1; 1; 1; 1; 1; 1; 1; 1; 1; 1Þ, ~m3 ¼
ð1; 1; 0; 1; 1; 0; 1; 1; 1; 1Þ, and ~m4 ¼ ð1; 1; 1; 1; 1; 1; 1; 1; 1; 1Þ.
The auditor calculates the autocorrelation function gj for
each sequence ~mj ¼ ðmj1; . . . ; mjMÞ, j ¼ 1; . . . ; K, as follows:
gjðiÞ ¼
PMÀi
k¼1 mjkmjkþi
M À i
; for i ¼ 0; . . . ; M À 1; j ¼ 1; . . . ; K: (9)
The auditor then calculates the relative difference between
gj and the ACF of the wireless channel fc as follows:
j ¼
XMÀ1
i¼0
gjðiÞ À fcðiÞ

fcðiÞ
: (10)
The relative difference j is then used as the decision statis-
tic to decide whether or not the packet loss over the jth hop
is caused by malicious drops. In particular, if j ! th, where
th is an error threshold, then Ad decides that there is mali-
cious packet drop over the hop. In this case, both ends of
the hop will be considered as suspects, i.e., either the trans-
mitter did not send out the packet or the receiver chose to
ignore the received packet. S may choose to exclude both
nodes from future packet transmissions, or alternatively,
apply a more extensive investigation to refine its detection.
For example, this can be done by combining the neighbor-
overhearing techniques [12] used in the reputation system.
By fusing the testimony from the neighbors of these two
nodes, Ad can pin-point the specific node that dropped the
packet. Once being detected, the malicious node will be
marked and excluded from the route to mitigate its damage.
The above detection process applies to one end-to-end
path. The detection for multiple paths can be performed as
multiple independent detections, one for each path. Although
the optimal error threshold that minimizes the detection error
is still an open problem, our simulations show that through
trial-and-error, one can easily find a good th that provides a
better detection accuracy than the optimal detection scheme
that utilizes only the pdf of the number of lost packets.
Public verifiability. After each detection, Ad is required to
publish the information it received from involved nodes,
i.e., ~bj, rðjÞ
, sðjÞ
, for j 2 PSD, so that a node can verify all cal-
culation has been performed correctly. Note that no knowl-
edge of the HLA secret key x is required in the verification
process. At the same time, because Ad has no knowledge of
x, there is no way for it to forge a valid HLA signature for
rðjÞ
. In other words, Ad cannot claim a misbehaving node to
be a normal one. Furthermore, the privacy-preserving prop-
erty of the scheme (see Theorem 4 in Section 4.3) ensures
that publishing the auditing information will not compro-
mise the confidentiality of the communication.
4.3 Security Analysis
We prove that the proposed scheme has the following secu-
rity properties.
Theorem 1. The verification of rðjÞ
and sðjÞ
, as specified in (8),
is correct, i.e., (8) must hold for a ð~cj; rðjÞ
, sðjÞ
Þ tuple that
is constructed according to the specification presented in
Section 4.2.3.
Proof. The correctness of (8) is shown as follows:
eðsðjÞ
; gÞ ¼ e
YM
i¼1;bji6¼0
s
cji
ji ; g
0
@
1
A
¼ e
YM
i¼1;bji6¼0
n
H2ðijjjÞuri
oxcji
; g
0
@
1
A
¼ e
YM
i¼1;bji6¼0
n
H2ðijjjÞuri
ocji
; g
0
@
1
A
x
¼ e
YM
i¼1;bji6¼0
H2ðijjjÞcji
ucjiri ; g
0
@
1
A
x
¼ e
YM
i¼1;bji6¼0
H2ðijjjÞcji
urðjÞ
; gx
0
@
1
A
¼ e
YM
i¼1;bji6¼0
H2ðijjjÞcji
urðjÞ
; v
0
@
1
A: (11)
So Theorem 1 holds. tu

Theorem 2. The construction specified in Section 4.2 is secure
under the collusion model defined in Section 3.2, i.e., an adver-
sary that does not receive a packet Pi cannot claim receiving
this packet in its ~bj by forging a HLA signature for a random
linear combination of the received packets, even if this adver-
sary colludes with any other malicious node in PSD.
Proof. For a given node nj, our construction essentially fol-
lows the BLS-signature-based HLA construction
described in [27]. Under the implicitly assumed condi-
tion of no collusion between attackers, the authors in [27]
proved that the construction is secure, i.e., no adversary
can forge a response to a random challenge if it does not
know the HLA signature of each packet in the linear
combination. So here, we only need to show that collu-
sion between malicious nodes does not give the attacker
more information about the HLA signature of the pack-
ets. This can be shown by observing the following novel
properties of our HLA construction:
1) For a packet Pi, the signature scheme specified in
(3) dictates that its HLA signature sji is not only
tied to the packet sequence number (i), but also
related to the node index (j) that is relaying the
packet. This means that for the same packet, each
hop on PSD is given a different HLA signature.
The verification scheme in (8) accounts for both i
and j. In the no-collusion case, by treating the con-
catenation of ðijjjÞ as a meta packet sequence
number, the security of our construction can be
proved in the same way as that in [27].
2) The way that the HLA signatures are distributed to
nodes on PSD, as specified in (4), dictates that an
upstream node nj cannot get a copy of the HLA sig-
nature sj0i of a downstream node nj0 , where
j j0
K, unless the downstream node nj0
receives the signature sj0i first on PSD and then
sends it through the covert channel to the upstream
node nj. Therefore, there is no way for a down-
stream malicious node to get any information on its
HLA signature if the upstream attacker drops the
packet. As a result, the secret information exchange
on the covert channel does not help the adversary
to get more information on its HLA signature than
the scenario where there is no collusion.
Combining the above arguments, Theorem 2 is
proved. tu
Theorem 3. The proposed scheme ensures that the packet-recep-
tion bitmap reported by a node in PSD is truthful.
The validity of Theorem 3 is straightforward, because
Theorem 2 guarantees that the node cannot understate its
packet loss information. At the same time, from our discus-
sion in Section 4.2.4, it is clear that a malicious node cannot
overstate its packet loss either. So a node must report its
actual packet reception information truthfully to Ad.
Theorem 4. Our HLA construction is publicly verifiable and pri-
vacy preserving, i.e., the auditor Ad does not require the secret
key of the HLA scheme to verify a node’s response. In addition,
Ad cannot determine the content of the packets transmitted
over PSD from the information submitted by nodes.
Proof. Public verifiability is clear from the construction of
the scheme. The privacy-preserving property is guaran-
teed by the application of the secure hash function H1.
More specifically, instead of directly computing the HLA
signature for a packet Pi, our construction computes the
signature for the image of the packet ri ¼ H1ðPiÞ. During
the auditing phase, Ad can collect a set of linear combina-
tions of ri’s. So it is possible for Ad to calculate ri’s by
solving a set of linear equations, if a sufficient number of
combinations are collected. Even if Ad can recover ri, it
should not be able to guess Pi because of H1’s resilience
to the pre-image attack. tu
4.4 Overhead Analysis
The proposed scheme requires relatively high computation
capability at the source, but incurs low communication and
storage overheads along the route, as explained below.
4.4.1 Computation Requirements
Most of the computation is done at the source node (for gen-
erating HLA signatures) and at the public auditor (for con-
ducting the detection process). We consider the public
auditor as a dedicated service provider that is not con-
strained by its computing capacity. So the computational
overhead should not be a factor limiting the application of
the algorithm at the public auditor. On the other hand, the
proposed algorithm requires the source node to generate K
HLA signatures for a K-hop path for each data packet. The
generation of HLA signatures is computationally expensive,
and may limit the applicability of the algorithm. We pro-
pose a block-based HLA signature and detection mecha-
nism in Section 5, whereby the processing is based on block
of packets rather than individual packets, to reduce this
computation overhead by multiple folds. We evaluate the
performance of the proposed mechanism by extensive sim-
ulations in Section 6.2.4.
4.4.2 Communication Overhead
The communication overhead for the setup phase is a one-
time cost, incurred when PSD is established. Here we mainly
focus on the recurring cost during the packet transmission
and auditing phases (there is no communication overhead
in the detection phase). For a transmitted packet Pi, S needs
to send one encrypted HLA signature and one MAC to each
intermediate node on PSD. Our HLA signature follows the
BLS scheme in [7]. So an HLA signature sij is 160-bit long. If
encrypted by DES, the encrypted signature ~sij is 192 bits in
length (a block in DES is 64-bit long, so the length of the
cipher text of DES is multiples of 64 bits). The MAC-related
hash function HMAC
key can be implemented in SHA-1 and has
a length of 160 bits. So for each packet, the per-hop commu-
nication overhead incurred by the proposed scheme in the
packet transmission phase is 192 þ 160 ¼ 352 bits, or
44 bytes. For a path of K intermediate hops, the total com-
munication overhead for transmitting a packet is 44K bytes.
For example, when K ¼ 10, the overhead is 440 bytes/
packet. For an IEEE 802.11 system, this is about 19 percent
of the maximum MSDU (2,304 bytes).

In the auditing phase, the auditor Ad sends a random chal-
lenge vector~cj to each node nj. Let each element in this vector
be a 32-bit integer. The challenge has a length of 4M bytes.
Based on our simulation in Section 6, M ¼ 50 is typically
enough to achieve good detection accuracy. So this means
each challenge can be delivered in one packet. Node nj replies
to the challenge with ~bj, rðjÞ
, and sðjÞ
. Among them, ~bj is an
M-bit bitmap. rðjÞ
is the linear combination of the SHA-1
image of the packets, so rðjÞ
also has a length of 160 bits. sðjÞ
is
an HLA signature of rðjÞ
, so it is also 160-bit long. Overall, the
reply from a node to Ad has a length of 320 þ M bits, which
can also be delivered in one packet.
4.4.3 Storage Overhead
During its operation, a node nj on PSD needs to store the
key keyj, the H1 hash image, and the associated HLA sig-
nature for each of the M most recently received packets.
Assuming encryptkey and decryptkey are based on DES, keyj
has a length of 56 bits. Let the hash function H1 be based
on SHA-1. So the H1 image of a packet is 160-bit long. The
HLA signature is based on BLS (Boneh-Lynn-Shacham)
scheme [7] and is 160-bit long. So in total the storage over-
head at nj is 320M þ 56 bits, or 40M þ 7 bytes. This storage
overhead is quite low. For example, when M 50, the
storage overhead at a node is less than 2 KB.
5 REDUCING COMPUTATION OVERHEAD:
BLOCK-BASED HLA SIGNATURE GENERATION
AND DETECTION
As discussed in Section 4.4.1, one major limitation of the
proposed baseline HLA detection algorithm is the high
computation overhead of the source node. In this section,
we proposed a block-based solution that can reduce this
overhead by multiple folds. The main idea is to make the
HLA signature scalable: instead of generating per-packet
HLA signatures, per-block HLA signatures will be gener-
ated, where a block consists of L 1 packets. Accordingly,
the detection will be extended to blocks, and each bit in the
packet-loss bitmap represents a block of packets rather than
a single packet. The details of this extension are elaborated
as follows.
In the Packet Transmission Phase, rather than generating
HLA signatures for every packet, now the signatures are
based on a block of packets. In particular, L consecutive
packets are deemed as one block. Accordingly, the stream
of packets is now considered as stream of blocks. Denote
the L packets in block i as Pi1; . . . ; PiL, respectively. The
source S generates block-HLA signatures for block i as
follows:
1) S computes ri ¼ H1ðPi1Þ. S then computes HLA sig-
natures of ri for node nj, say sji, where j ¼ 1; . . . ; K,
according to (3).
2) For each node nj, S generates L random numbers
k
ð1Þ
ji ; . . . ; k
ðLÞ
ji 2 Zp, such that
PL
l¼1 k
ðlÞ
ji ¼ sji. This could
be done, e.g., but first generating L À 1 arbitrary
numbers and then make the Lth number equal to
sji À
PLÀ1
l¼1 k
ðlÞ
ji .
3) For packet Pil, S assigns k
ðlÞ
ji ’s, j ¼ 1; . . . ; K, as the
block-HLA signatures for node 1; . . . ; K, respec-
tively. These signatures are then transmitted with
packet Pil by following the one-way chained
encryption and decryption scheme described in (4)
through (6).
In line with the above block-HLA signatures, a node nj is
able to compute sji ¼
PL
l¼1 k
ðlÞ
ji if and only if it receives all L
packets of block i. Node nj stores ri and sji in its proof-of-
reception database as the proof for block i.
Auditing is now based on blocks. In particular, at node
nj, suppose the sequence number of the blocks recorded in
the proof-of-reception database are B1; . . . ; BM . Based on
the information in the database, node nj generates a block-
reception bitmap ~bj ¼ ðbj1; . . . ; bjMÞ, where bji ¼ 1 if and
only if all L packets in block Bi has been received by nj, and
bji ¼ 0 otherwise. Except the above, Ad still follows the
same algorithm in Section 4.2.3 to submit random challenge,
receive response, and verify the truthfulness of the reported
block-reception bitmap.
In the detection phase, the ACF of the wireless channel
needs to be coarsened such that one unit of lag represents L
consecutive packets. This could be done by first coarsening
the packet reception bitmap observed in the training phase
using blocks: L consecutive 1’s are mapped to a 1 in the
blocked-based bitmap, otherwise a 0 will be mapped. The
ACF of the coarsened wireless channel is then compared
with the ACF of the block-reception bitmap reported by
each node to detect possible malicious packet drops.
From the above description, it is clear that the block-
based HLA signature and detection mechanism can in gen-
eral reduce the computation overhead by L folds. However,
the coarser representation of lost packets makes it difficult
to accurately capture the correlation between them. For
example, even with a small block size, say L ¼ 2, it is not
possible to tell whether a block loss is due to the loss of one
packet or both packets in the block, which correspond to
very different packet-loss correlations. Therefore, it is
expected that the reduced computational overhead comes
at the cost of less detection accuracy. We study the perfor-
mance of the block-based algorithm in Section 6.2.4.
6 PERFORMANCE EVALUATION
6.1 Simulation Setup
In this section, we compare the detection accuracy
achieved by the proposed algorithm with the optimal max-
imum likelihood algorithm, which only utilizes the distri-
bution of the number of lost packets. For given packet-loss
bitmaps, the detection on different hops is conducted sepa-
rately. So, we only need to simulate the detection of one
hop to evaluate the performance of a given algorithm. We
assume packets are transmitted continuously over this hop,
i.e., a saturated traffic environment. We assume channel
fluctuations for this hop follow the Gilbert-Elliot model,
with the transition probabilities from good to bad and
from bad to good given by PGB and PBG, respectively. We
consider two types of malicious packet dropping: random
dropping and selective dropping. In the random dropping
attack, a packet is dropped at the malicious node with

probability PM . In the selective dropping attack, the adver-
sary drops packets of certain sequence numbers. In our
simulations, this is done by dropping the middle N of the
M most recently received packets, i.e., setting the N bits in
the middle of the packet-loss bitmap to 0 (if a packet in
these positions is dropped due to link errors, then the set
of 0’s extends to an extra bit in the middle). PM and N are
simulation parameters that describe the selectivity of the
attack. In both cases, we let th ¼ 10% for the proposed
algorithm.
We are interested in the following three performance
metrics: probability of false alarm (Pfa), probability of miss-
detection (Pmd), and the overall detection-error probability
(Perror). We collect these statistics as follows. In each run, we
first simulate 1,000 independently generated packet-loss bit-
maps for the hop, where packet losses are caused by link
errors only. We execute our detection algorithm over these
packet-loss bitmaps and collect the number of cases where
the algorithm decides that an attacker is present. Let this
number be Ifa. Pfa of this run is calculated as Pfa ¼ Ifa=
1;000. We then simulate another 1,000 independently gener-
ated packet-loss bitmaps, where losses are now caused by
both link errors and malicious drops. Let the number of
cases where the detection algorithm rules that an attacker is
not present be Imd. Pmd of the underlying run is given by
Pmd ¼ Imd=1;000. Perror is given by Perror ¼ ðIfa þ ImdÞ=
2; 000. The above simulation is repeated 30 times, and the
mean and 95 percent confidence interval are computed for
the various performance metrics.
6.2 Results
6.2.1 Random Packet Dropping
The detection accuracy is shown in Fig. 4 as a function of the
malicious random-drop rate PM. In each subfigure, there are
two sets of curves, representing the proposed algorithm and
the optimal ML scheme, respectively. In each set of curves,
the one in the middle represents the mean, and the other two
represent the 95 percent confidence interval. In general, the
detection accuracy of both algorithms improves with PM (i.e.,
the detection error decreases with PM). This is not surprising,
because malicious packet drops become more statistically dis-
tinguishable as the attacker starts to drop more packets. In
addition, this figure shows that for th ¼ 10%, the proposed
algorithm provides slightly higher false-alarm rate (subfigure
(c)) but significantly lower miss-detection probability
(subfigure (b)) than the ML scheme. A low miss-detection
probability is very desirable in our context, because it means a
malicious node can be detected with a higher probability. The
slightly higher false-alarm rate should not be a problem,
because a false alarm can be easily recognized and fixed in the
post-detection investigation phase. Most importantly, the
overall detection-error probability of the proposed scheme is
lower than that of the ML scheme (subfigure (a)). We are espe-
cially interested in the regime when PM is comparable to the
average packet loss rate due to link errors, given by
PGB
PGBþPBG
¼ 0:01
0:01þ0:5 % 0:02. This regime represents the scenario
in which the attacker hides its drops in the background of link
errors by mimicking the channel-related loss rate. In this case,
the ML scheme cannot correctly differentiate between link
errors and malicious drops. For example, when PM ¼ 0:01,
the ML scheme results in Pmd ¼ 80% and Pfa ¼ 23%. This is
close to arbitrarily ruling that every packet loss is due to link
error only, leading to an overall detection-error rate of 50 per-
cent (see subfigure (a)). Our proposed algorithm, on the other
hand, achieves a much better detection accuracy, because its
Pmd and Pfa are both lower than those under the ML scheme.
As a result, when PM ¼ 0:01, the total detection-error rate of
the proposed algorithm is about 35 percent. When PM is
increased to 0.04, Perror of the proposed scheme reduces to
only 20 percent, which is roughly half of the error rate of the
ML scheme at the same PM. Remembering that the detection-
error rate of the ML scheme is the lowest among all detection
schemes that only utilize the distribution of the number of
lost packets, the lower detection-error rate of the proposed
scheme shows that exploiting the correlation between lost
packets helps in identifying the real cause of packet drops
more accurately. The effect of exploiting the correlation is
especially visible when the malicious packet-drop rate is com-
parable with the link error rate. Meanwhile, we also note that
the 95 percent confidence interval of the proposed scheme is
wider than that of the ML scheme. This is because the decision
variable j in the proposed scheme is a second-order function
of the random packet loss process, while the decision variable
in the ML scheme (i.e., number of lost packets) is a first order
function of the same packet loss process. As a result, the deci-
sion variable of the proposed scheme possesses more ran-
domness than that of the ML scheme, as reflected by the
wider 95 percent confidence interval.
In Fig. 5, we plot the detection accuracy as a function
of the size of the packet-loss bitmap ðMÞ. It can be
Fig. 4. Detection accuracy versus PM (random packet-drop case).

observed that Perror for the proposed scheme decreases
with M. However, as M becomes sufficiently large, e.g.,
M ¼ 30 in our case, a further increase in the size of the
bitmap does not lead to additional improvement in the
detection accuracy. This can be explained by noting that
the two-state Markovian GE channel model has a short-
range dependence, i.e., the correlation between two
points of the fluctuation process decays rapidly with the
increase in the separation between these points. This
short-range dependence is reflected in an exponentially
decaying autocorrelation function for the channel. As a
result, a good estimation of the autocorrelation function
can be derived as long as M is long enough to cover the
function’s short tail. This phenomenon implies that a
node does not need to maintain a large packet-reception
database in order to achieve a good detection accuracy
under the proposed scheme. It also explains the low stor-
age overhead incurred by our scheme.
The detection accuracy is plotted in Fig. 6 as a function of
the channel state transition rate PGB. It can be observed
from this figure that Perror for both algorithms increases
with PGB. This is not surprising because at its initial point of
PGB ¼ 0:01, the expected link error rate is about 0.02, which
is much smaller than the malicious packet drop rate of
PM ¼ 0:1. So it is relatively easy to differentiate between the
case where packet drops are caused by link errors only and
the one where such drops are caused by the combined effect
of link errors and malicious drops. As PGB increases, the
link error probability approaches PM, making the statistical
separation of the two cases harder. As a result, the detection
error increases with PGB. For all values of PGB in this figure,
the proposed algorithm always achieves significantly lower
detection-error probability than the ML scheme.
6.2.2 Selective Packet Dropping
The detection error as a function of the number of mali-
ciously dropped packets is shown in Fig. 7. At the low end
of the x-axis, maliciously dropped packets account for only
1=50 ¼ 2% of the total packets in the packet-loss bitmap.
This is identical to the link error rate of 0.02, assumed in the
simulation. Similar performance trends can be observed to
the case of the random packet dropping. Fewer detection
errors are made by both algorithms when more packets are
maliciously dropped. In all the simulated cases, the pro-
posed algorithm can detect the actual cause of the packet
drop more accurately than the ML scheme, especially when
the number of maliciously dropped packets is small. When
the number of maliciously dropped packets is significantly
higher than that caused by link errors (greater than four
packets in our simulation), the two algorithms achieve com-
parable detection accuracy. In this scenario, it may be wise
to use the conventional ML scheme due to its simplicity
(e.g., no need to enforce truthful reports from intermediate
nodes, etc.).
The detection errors are plotted in Fig. 8 as a function of
the size of the packet-loss bitmap (M). To conduct a fair
comparison, as we increase M, we also increase the number
of maliciously dropped packets, so as to maintain a mali-
cious packet-dropping rate of 10 percent. It can be observed
that a small M is enough to achieve good detection accuracy
Fig. 5. Detection accuracy versus M (random packet-drop case).
Fig. 6. Detection accuracy versus PGB (random packet-drop case).

under the proposed scheme, due to the short-range depen-
dence property of the channel.
In Fig. 9, the detection errors are plotted as a function of
the channel state transition probability PGB. Similar trends
are observed to those in the random packet dropping case,
i.e., the algorithms make more detection errors when the
link error rate approaches the malicious packet-drop rate.
Once again, the proposed algorithm consistently outper-
forms the ML scheme in all the tested cases.
6.2.3 Dropping of Control Packets
Our simulations so far have not made any application-
semantic (use case) assumption on the dropped packets. In
reality, however, because these packets are usually used for
control purposes, the loss of these packets may generate sig-
nificant impacts on the transmission of other (i.e., data)
packets. In this series of simulations, we evaluate how the
correlation between the control and data packets affects the
performance of the proposed scheme. In particular, we con-
sider a multi-hop cognitive radio network, where control
packets are exchanged over an end-to-end path to maintain
channel synchronization between consecutive hops. A con-
trol packet contains the channel id that the two ends of a
hop should tune to. The exchange of these packets could be
end-to-end (i.e., the entire path operates over a channel
commonly available to all hops) or local (the path consists
of several segments, each of which operates over a different
channel available only to the hops of that segment). In either
case, the drop of a control packet will disrupt the frequency
synchronization of a hop, so that data packets transmitted
over the hop will be lost until the frequency synchronization
is resumed by the next valid exchange of control packets. In
this simulation, we assume that a random number of ldata
data packets are sent between two consecutive control pack-
ets. We assume that ldata follows geometric distribution with
mean parameter Ldata (i.e., the continuous idle time of a
channel is exponentially distributed, as commonly assumed
in the cognitive radio literature). In addition, we also
assume that an attacker uniformly randomly drops a certain
percentage of the control packets. This percentage and Ldata
are parameters varied in the simulation. Finally, we use the
same Gilbert-Elliot model as previous simulations to gener-
ate the wireless channel.
Fig. 10a plots detection accuracy as a function of the con-
trol packet dropping rate. It can be observed that under the
correlated control and data packet losses, the proposed
scheme achieves significantly better detection accuracy than
the ML scheme for all tested control packet dropping rates, a
similar trend to those observed when such correlation is not
considered (e.g., see the random packet drops in Fig. 4a).
Meanwhile, it can also be observed from Figs. 10a and 4a
that, compared with the results of uncorrelated packet drops,
the detection-error probability under correlated packet
losses is in general smaller, indicating that the correlation
between control and data packet losses may help to improve
the detection accuracy. This is true for both schemes. This is
because such correlation further amplifies the distinction
between the statistics of the wireless channel and the mali-
cious packet drops, making the detection easier.
Fig. 7. Detection accuracy versus number of maliciously dropped packets (selective packet-drop case).
Fig. 8. Detection accuracy versus M (selective packet-drop case).

We study the detection accuracy as a function of aver-
age number of data packets transmitted between two
consecutive control packets (Ldata) in Fig. 10b. It can be
observed that this parameter has little impact on the ML
scheme, because under a given control packet loss rate,
the overall (control+data) packet loss rate does not
change with Ldata. ML scheme relies on detecting the
number of lost packets, and thus is not affected by Ldata.
In contrast, the proposed scheme may benefit from a
larger Ldata: even if the average detection accuracy does
not improve much, its 95 percent confidence interval
shrinks significantly with Ldata. This is explained by
comparing the packet loss patterns of the wireless chan-
nel and the correlated packet drops. Specifically, on aver-
age two packets are lost in a row in a wireless link error
(this is given by 1=ð1 À PBGÞ ¼ 1=0:5 ¼ 2), while on aver-
age Ldata packets are lost under each malicious drop.
Therefore, the distinction between the above two packet-
loss patterns increases with Ldata. The proposed scheme
exploits such a distinction in patterns and thus can bene-
fit from larger Ldata.
In Fig. 10c we plot the detection-error probability as a
function of channel state transition parameter PGB. It can
be observed that the detection accuracy deteriorates with
the increase of PGB, because it becomes more and more
difficult to decide the actual source of packet loss when
link error rate approaches the malicious packet dropping
rate. Once again, the proposed scheme consistently
achieves higher detection accuracy than ML scheme in
all simulated cases.
6.2.4 Block-Based Detection
In this series of simulations, we study the detection accuracy
of block-based algorithms as a function of block size.
Fig. 11a plots the detection accuracy for random packet
drops under two packet drop probabilities: high (PM ¼ 0:08)
and low (PM ¼ 0:01). The performance of the ML scheme is
also plotted in the same figure for comparison. In general, it
shows that for both cases the detection error increases with
the block size. This is expected, as a larger block size hides
more details of packet losses, and therefore makes the actual
correlation of lost packets more difficult to calculate. Mean-
while, the benefits of blocked-based algorithm is also
observed: it is able to trade computation complexity for bet-
ter detection accuracy. For example, under low packet drop-
ping rate, it shows that the block-based algorithm can
reduce the computation overhead of the baseline HLA
detection by 10 folds, and still be able to achieve better detec-
tion accuracy than the ML scheme. At high packet dropping
rate, the block-based algorithm can achieves a 4Â computa-
tion overhead reduction, while still achieving slightly better
detection accuracy than the ML scheme.
Fig. 11b plots the detection accuracy for random packet
drops under two packet sampling methods. In the first
method, we fix the total number of packets used in the sam-
ple, and therefore the number of sampled blocks varies
with the block size. In the second method, we fix the num-
ber of sampled blocks, but the number of sampled packets
changes with the block size. Note that the amount of com-
putation required to compute the block-based signatures
decreases with the block size in the first method, but
Fig. 9. Detection accuracy versus PGB (selective packet-drop case).
Fig. 10. Detection accuracy for control packet drops.

remains the same in the second method. The second method
does not reduce the amount of computation, rather, it
reduces the intensity of the computation by distributing it
over a larger time interval (more packets). From this figure,
it can be observed that under both methods, the detection
accuracy deteriorates with the block size, but the deteriora-
tion is more severe under the first method. This is not sur-
prising, because in method one the block-reception bitmap
becomes shorter with the increase of block size, and there-
fore the computed ACF is less accurate, due to the insuffi-
cient sample size. Meanwhile, it can be observed that, at a
detection accuracy comparable to that of the ML scheme,
the second method achieves 4Â computation overhead
reduction over the baseline algorithm, whereas the saving
of the first method is only 2Â. This observation suggests
that, to maintain a good detection accuracy, block-based
algorithm may be better used as a way to reduce the compu-
tation intensity of the source node, rather than a way to
reduce the absolute amount of the computation.
Fig. 11c plots the detection accuracy for selective packet
drops under two packet dropping rates: high (three out of
every 50 packets are dropped) and low (one out of every
50 packets is dropped). A similar trend to Fig. 11a can be
observed. This observation suggests that the property–
block-based algorithm can trade computation complexity
for detection accuracy—is universal (i.e., holds under vari-
ous attack models).
7 CONCLUSIONS
In this paper, we showed that compared with conventional
detection algorithms that utilize only the distribution of the
number of lost packets, exploiting the correlation between
lost packets significantly improves the accuracy in detecting
malicious packet drops. Such improvement is especially vis-
ible when the number of maliciously dropped packets is
comparable with those caused by link errors. To correctly
calculate the correlation between lost packets, it is critical to
acquire truthful packet-loss information at individual
nodes. We developed an HLA-based public auditing archi-
tecture that ensures truthful packet-loss reporting by indi-
vidual nodes. This architecture is collusion proof, requires
relatively high computational capacity at the source node,
but incurs low communication and storage overheads over
the route. To reduce the computation overhead of the
baseline construction, a packet-block-based mechanism was
also proposed, which allows one to trade detection accuracy
for lower computation complexity.
Some open issues remain to be explored in our future
work. First, the proposed mechanisms are limited to static or
quasi-static wireless ad hoc networks. Frequent changes on
topology and link characteristics have not been considered.
Extension to highly mobile environment will be studied in
our future work. In addition, in this paper we have assumed
that source and destination are truthful in following the
established protocol because delivering packets end-to-end
is in their interest. Misbehaving source and destination will
be pursued in our future research. Moreover, in this paper,
as a proof of concept, we mainly focused on showing the fea-
sibility of the proposed cypto-primitives and how second-
order statistics of packet loss can be utilized to improve
detection accuracy. As a first step in this direction, our analy-
sis mainly emphasize the fundamental features of the prob-
lem, such as the untruthfulness nature of the attackers, the
public verifiability of proofs, the privacy-preserving require-
ment for the auditing process, and the randomness of wire-
less channels and packet losses, but ignore the particular
behavior of various protocols that may be used at different
layers of the protocol stack. The implementation and optimi-
zation of the proposed mechanism under various particular
protocols will be considered in our future studies.
ACKNOWLEDGMENTS
The work of T. Shu was supported in part by US National
Science Foundation (NSF) under Grant CNS-1343156. The
work of M. Krunz was supported by NSF (Grants CNS-
1016943 and IIP-1265960) and by ARO (Grant W911NF-13-
1-0302). Any opinions, findings, conclusions, or recommen-
dations expressed in this paper are those of the author(s)
and do not necessarily reflect the views of NSF or ARO. A
preliminary version of this paper was presented in ACM
WiSec 2012 Conference, Tucson, AZ, April 2012.
REFERENCES
[1] J. N. Arauz, “802.11 Markov channel modeling,” Ph.D. dissertation,
School Inform. Sci., Univ. Pittsburgh, Pittsburgh, PA, USA, 2004.
[2] C. Ateniese, R. Burns, R. Curtmola, J. Herring, L. Kissner, Z.
Peterson, and D. Song, “Provable data possession at untrusted
stores,” in Proc. ACM Conf. Comput. and Commun. Secur., Oct. 2007,
pp. 598–610.
Fig. 11. Detection accuracy of block-based algorithms.

[3] G. Ateniese, S. Kamara, and J. Katz, “Proofs of storage from homo-
morphic identification protocols,” in Proc. Int. Conf. Theory Appl.
Cryptol. Inf. Security, 2009, pp. 319–333.
[4] B. Awerbuch, R. Curtmola, D. Holmer, C. Nita-Rotaru, and H.
Rubens, “ODSBR: An on-demand secure byzantine resilient rout-
ing protocol for wireless ad hoc networks,” ACM Trans. Inform.
Syst. Security, vol. 10, no. 4, pp. 1–35, 2008.
[5] B. Awerbuch, R. Curtmola, D. Holmer, C. Nita-Rotaru, and H.
Rubens, “ODSBR: An on-demand secure byzantine resilient rout-
ing protocol for wireless ad hoc networks,” ACM Trans. Inf. Syst.
Secur., vol. 10, no. 4, pp. 11–35, 2008.
[6] K. Balakrishnan, J. Deng, and P. K. Varshney, “TWOACK: Pre-
venting selfishness in mobile ad hoc networks,” in Proc. IEEE
Wireless Commun. Netw. Conf., 2005, pp. 2137–2142.
[7] D. Boneh, B. Lynn, and H. Shacham, “Short signatures from the
weil pairing,” J. Cryptol., vol. 17, no. 4, pp. 297–319, Sep. 2004.
[8] S. Buchegger and J. Y. L. Boudec, “Performance analysis of the
confidant protocol (cooperation of nodes: Fairness in dynamic ad-
hoc networks),” in Proc. 3rd ACM Int. Symp. Mobile Ad Hoc Netw.
Comput. Conf., 2002, pp. 226–236.
[9] L. Buttyan and J. P. Hubaux, “Stimulating cooperation in self-
organizing mobile ad hoc networks,” ACM/Kluwer Mobile Netw.
Appl., vol. 8, no. 5, pp. 579–592, Oct. 2003.
[10] J. Crowcroft, R. Gibbens, F. Kelly, and S. Ostring, “Modelling
incentives for collaboration in mobile ad hoc networks,” pre-
sented at the First Workshop Modeling Optimization Mobile, Ad
Hoc Wireless Netw., Sophia Antipolis, France, 2003.
[11] J. Eriksson, M. Faloutsos, and S. Krishnamurthy, “Routing amid
colluding attackers,” in Proc. IEEE Int. Conf. Netw. Protocols, 2007,
pp. 184–193.
[12] W. Galuba, P. Papadimitratos, M. Poturalski, K. Aberer, Z.
Despotovic, and W. Kellerer, “Castor: Scalable secure routing for
ad hoc networks,” in Proc. IEEE INFOCOM, Mar. 2010, pp. 1 –9.
[13] T. Hayajneh, P. Krishnamurthy, D. Tipper, and T. Kim, “Detecting
malicious packet dropping in the presence of collisions and chan-
nel errors in wireless ad hoc networks,” in Proc. IEEE Int. Conf.
Commun., 2009, pp. 1062–1067.
[14] Q. He, D. Wu, and P. Khosla, “Sori: A secure and objective reputa-
tion-based incentive scheme for ad hoc networks,” in Proc. IEEE
Wireless Commun. Netw. Conf., 2004, pp. 825–830.
[15] D. B. Johnson, D. A. Maltz, and J. Broch, “DSR: The dynamic
source routing protocol for multi-hop wireless ad hoc networks,”
in Ad Hoc Networking. Reading, MA, USA: Addison-Wesley, 2001,
ch. 5, pp. 139–172.
[16] W. Kozma Jr. and L. Lazos, “Dealing with liars: Misbehavior
identification via Renyi-Ulam games,” presented at the Int.
ICST Conf. Security Privacy in Commun. Networks, Athens,
Greece, 2009.
[17] W. Kozma Jr., and L. Lazos, “REAct: Resource-efficient
accountability for node misbehavior in ad hoc networks based
on random audits,” in Proc. ACM Conf. Wireless Netw. Secur.,
2009, pp. 103–110.
[18] K. Liu, J. Deng, P. Varshney, and K. Balakrishnan, “An acknowl-
edgement-based approach for the detection of routing misbehav-
ior in MANETs,” IEEE Trans. Mobile Comput., vol. 6, no. 5,
pp. 536–550, May 2006.
[19] Y. Liu and Y. R. Yang, “Reputation propagation and agreement in
mobile ad-hoc networks,” in Proc. IEEE WCNC Conf., 2003,
pp. 1510–1515.
[20] S. Marti, T. J. Giuli, K. Lai, and M. Baker, “Mitigating routing mis-
behavior in mobile ad hoc networks,” in Proc. ACM MobiCom
Conf., 2000, pp. 255–265.
[21] G. Noubir and G. Lin, “Low-power DoS attacks in data wireles
lans and countermeasures,” ACM SIGMOBILE Mobile Comput.
Commun. Rev., vol. 7, no. 3, pp. 29–30, Jul. 2003.
[22] V. N. Padmanabhan and D. R. Simon, “Secure traceroute to detect
faulty or malicious routing,” in Proc. ACM SIGCOMM Conf., 2003,
pp. 77–82.
[23] P. Papadimitratos and Z. Haas, “Secure message transmission in
mobile ad hoc networks,” Ad Hoc Netw., vol. 1, no. 1, pp. 193–209,
2003.
[24] A. Proano and L. Lazos, “Selective jamming attacks in wireless
networks,” in Proc. IEEE ICC Conf., 2010, pp. 1–6.
[25] A. Proano and L. Lazos, “Packet-hiding methods for preventing
selective jamming attacks,” IEEE Trans. Depend. Secure Comput.,
vol. 9, no. 1, pp. 101–114, Jan./Feb. 2012.
[26] R. Rao and G. Kesidis, “Detecting malicious packet dropping
using statistically regular traffic patterns in multihop wireless net-
works that are not bandwidth limited,” in Proc. IEEE GLOBECOM
Conf., 2003, pp. 2957–2961.
[27] H. Shacham and B. Waters, “Compact proofs of retrievability,” in
Proc. Int. Conf. Theory Appl. Cryptol. Inf. Secur., Dec. 2008, pp. 90–
107.
[28] T. Shu, M. Krunz, and S. Liu, “Secure data collection in wireless
sensor networks using randomized dispersive routes,” IEEE
Trans. Mobile Comput., vol. 9, no. 7, pp. 941–954, Jul. 2010.
[29] T. Shu, S. Liu, and M. Krunz, “Secure data collection in wireless
sensor networks using randomized dispersive routes,” in Proc.
IEEE INFOCOM Conf., 2009, pp. 2846–2850.
[30] C. Wang, Q. Wang, K. Ren, and W. Lou, “Privacy-preserving pub-
lic auditing for data storage security in cloud computing,” in Proc.
IEEE INFOCOM Conf., Mar. 2010, pp. 1–9.
[31] W. Xu, W. Trappe, Y. Zhang, and T. Wood, “The feasibility of
launching and detecting jamming attacks in wireless networks,”
in Proc. ACM MobiHoc Conf., 2005, pp. 46–57.
[32] Y. Xue and K. Nahrstedt, “Providing fault-tolerant ad-hoc routing
service in adversarial environments,” Wireless Pers. Commun., Spe-
cial Issue Secur. Next Generation Commun., vol. 29, no. 3, pp. 367–
388, 2004.
[33] Y. Zhang, L. Lazos, and W. Kozma, “AMD: Audit-based misbe-
havior detection in wireless ad hoc networks,” IEEE Trans. Mobile
Comput., PrePrint, Vol. 99, published online on 6 Sept. 2013.
[34] S. Zhong, J. Chen, and Y. R. Yang, “Sprite: A simple cheat-proof,
credit-based system for mobile ad-hoc networks,” in Proc. IEEE
INFOCOM Conf., 2003, pp. 1987–1997.

Tao Shu received the BS and MS degrees in
electronic engineering from the South China Uni-
versity of Technology, Guangzhou, China, in
1996 and 1999, respectively, and the PhD degree
in communication and information systems from
Tsinghua University, Beijing, China, in 2003. He
received the PhD degree in electrical and com-
puter engineering from the University of Arizona
in December 2010. He was a senior engineer at
Qualcomm Inc. at San Jose, CA, from December
2010 to July 2011. He is currently an assistant
professor at the Department of Computer Science and Engineering,
Oakland University. His research aims at addressing the security, pri-
vacy, and performance issues in wireless networking systems, with
strong emphasis on system architecture, protocol design, and perfor-
mance modeling and optimization.
Marwan Krunz received the PhD degree in elec-
trical engineering from Michigan State University
in 1995. He is a professor of electrical and com-
puter engineering at the University of Arizona. He
holds a joint (courtesy) appointment at the same
rank at the Department of Computer Science. He
is the site co-director of Broadband Wireless
Access and Applications Center (BWAC), a
recently inaugurated US National Science Foun-
dation (NSF)/industry IUCRC center that includes
ﬁve universities and 20þ industry members.
From 2008 to 2014, he was the UA site director of the NSF IUCRC
“Connection One” center. He joined the University of Arizona in January
1997, after a brief postdoctoral stint at the University of Maryland. In
2010, he was a visiting chair of excellence at the University of Carlos III
de Madrid, and concurrently a visiting researcher at Institute IMDEA Net-
works. In 2011, he received the Fulbright Senior Expert Award, through
which he visited with the University of Jordan, King Abdullah II School of
Information Technology. He previously held other visiting research posi-
tions at INRIA-Sophia Antipolis, HP Labs - Palo Alto, University of Paris
VI, University of Paris V, and US West Advanced Technologies. His
research interests lie in the areas of wireless communications and net-
working, with emphasis on resource management, adaptive protocols,
and security issues. Recently, he has been involved in projects related
to cognitive radios and dynamic spectrum access, wireless security,
power-controlled protocols for wireless networks, multichannel MIMO
systems, secure satellite communications, energy management in
solar-powered WSNs, and full-duplex communications. He has pub-
lished more than 200 journal articles. He received the 2012 IEEE Com-
munications Society TCCC Outstanding Service Award. He received the
US NSF CAREER award in 1998. He currently serves on the editorial
board for the IEEE Transactions on Network and Service Management.
Previously, he served on the editorial boards for the IEEE/ACM Transac-
tions on Networking, the IEEE Transactions on Mobile Computing, the
Computer Communications Journal, and the IEEE Communications
Interactive Magazine. He was a guest coeditor for special issues in IEEE
Micro and IEEE Communications Magazines. He was the general
cochair for the ACM WiSec 2012 Conference, and served as a TPC
chair for various international conferences, including INFOCOM’04,
SECON’05, and WoWMoM’06. He has served and continues to serve
on the executive and technical program committees of numerous inter-
national conferences, and on the panels of several funding agencies. He
was the keynote speaker, an invited panelist, and a tutorial presenter at
various international conferences. He is a fellow of the IEEE, an Arizona
Engineering Faculty Fellow (2011-2014), and an IEEE Communications
Society Distinguished lecturer (2013 and 2014).
For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.

IEEE BE-BTECH NS2 PROJECT@ DREAMWEB TECHNO SOLUTION

More Related Content

What's hot (17)

Viewers also liked (15)

Similar to IEEE BE-BTECH NS2 PROJECT@ DREAMWEB TECHNO SOLUTION (20)

Recently uploaded (20)

IEEE BE-BTECH NS2 PROJECT@ DREAMWEB TECHNO SOLUTION