enhancing ieee 802.11 random backoff in selfish environments

1806 IEEE TRANSACTIONS ON VEHICULAR TECHNOLOGY, VOL. 57, NO. 3, MAY 2008

Enhancing IEEE 802.11 Random Backoffin Selfish Environments

Lei Guang, Student Member, IEEE, Chadi M. Assi, Member, IEEE, and Abderrahim Benslimane, Member, IEEE

Abstract—Wireless access protocols currently deployed inmobile ad hoc networks use distributed contention resolutionmechanisms for sharing the wireless channel. In such an envi-ronment, selfish hosts that fail to adhere to the medium accesscontrol (MAC) protocol may obtain an unfair share of the channelbandwidth at the expense of performance degradation of well-behaved hosts. We present a novel access method, called pre-dictable random backoff (PRB), that is capable of mitigating themisbehavior of selfish hosts, particularly hosts that deliberatelydo not respect the random deferment of the transmission of theirpackets. PRB is based on minor modifications of the IEEE 802.11binary exponential backoff (BEB) and forces each node to generatea predictable backoff interval. The key idea is to adjust, in apredictable manner, the lower bound of the contention windowto enhance the per-station fairness in selfish environments. Hoststhat do not follow the operation of PRB are therefore easilydetected and isolated. We present an accurate analytical modelto compute the system throughput using a 3-D Markov chain.We evaluate the performance of PRB under the normal case andin the presence of selfish hosts. Our results show that PRB andBEB similarly perform in the former case. Selfish hosts, however,achieve substantially higher throughput than well-behaved hostsunder BEB. PRB, on the other hand, can effectively enhanceIEEE 802.11 BEB by mitigating the impacts of these MAC selfishmisbehaviors and guarantee a fair share of the wireless channelfor well-behaved hosts.

Index Terms—Ad hoc networks, IEEE 802.11, medium accesscontrol (MAC), performance evaluation, selfish misbehavior.

I. INTRODUCTION

R ELIABLE communication in ad hoc networks dependson the inherent trust among nodes. Trust means that

nodes need to fully cooperate with each other to ensure cor-rect route establishment mechanisms, protection of routinginformation, and security of packet forwarding [5], [14]–[16],[19], [20], [22]. However, this trust might be abused by ad-versaries to carry out security breaches through compromisednodes. The traditional approach to providing network securityis built on cryptography-based authentication. However, thisis not sufficient to solve the problems arising from new node

Manuscript received September 6, 2006; revised June 1, 2007 and August13, 2007. The review of this paper was coordinated by Prof. Y.-C. Tseng.

L. Guang is with Voice & Data Systems Inc., Montreal, QC H3G 1T7,Canada (e-mail: [email protected]).

C. M. Assi is with the Concordia Institute for Information Systems En-gineering, Concordia University, Montreal, QC H3G 1M8, Canada (e-mail:[email protected]).

A. Benslimane is with the Laboratoire Informatique d’Avignon, Universitéd’Avignon, 84914 Avignon Cédex 9, France (e-mail: [email protected]).

Color versions of one or more of the figures in this paper are available onlineat http://ieeexplore.ieee.org.

Digital Object Identifier 10.1109/TVT.2007.909291

misbehaviors in mobile ad hoc networks (MANETs). Hence,securing MANETs against medium access control (MAC) layermisbehavior has become a major challenge in the researchcommunity.

Host misbehaviors in MANETs can be classified into thefollowing two categories: 1) selfish misbehavior [7], [18] and2) malicious misbehavior [2], [12]. A selfish host can deliber-ately misuse the MAC protocol to gain more network resourcesthan well-behaved hosts. The node can benefit from this behav-ior through the following three scenarios: 1) obtaining a largeportion of channel capacity (hence, an improved throughput);2) reduced power consumption; and (3) improved quality ofservice, e.g., low network latency. For example, IEEE 802.11requires hosts that compete for the channel to wait for backoffinterval [18] before any transmissions. A selfish host maychoose to wait for a smaller backoff interval, thereby increasingits chance of accessing the channel and, hence, reducing thethroughput share received by well-behaved stations. Kyasanurand Vaidya [18] showed that such selfish misbehavior can seri-ously degrade the performance of the network, and accordingly,they proposed some modifications for the protocol (e.g., byallowing the receiver to assign backoff values rather than thesender) to detect and penalize misbehaving nodes. Similarly,Raya et al. [21] addressed the same problem and proposed asystem, called DOMINO, to detect greedy misbehaviors suchas backoff manipulations in IEEE 802.11.

Alternatively, malicious misbehaviors primarily aim at dis-rupting the normal operation of the network. This includescolluding adversaries that continuously send data to each otherto deplete the channel capacity in their vicinity (i.e., causinga denial-of-service attack) and, hence, prevent other legitimateusers from communicating [23].

To mitigate the impact of selfish nodes on the network perfor-mance, particularly those smart nodes [10], we proposed a newadaptive and predictable algorithm, called predictable randombackoff (PRB), that is based on minor modifications of theIEEE 802.11 binary exponential backoff (BEB), as explainedin the subsequent sections. The motivation of PRB is theenhancement of BEB performance for well-behaved stationsin a selfish environment and the facilitation of the detectionprocedures against potential selfish stations. Another majorcontribution of this paper is the development of an accurateanalytical model for the evaluation of the performance of PRBby employing a 3-D Markov chain analysis. We consider in ourstudy the performance of PRB under the following two cases:1) the normal behavior and 2) the selfish behavior. Further-more, we focus, throughout the paper, on single-hop wirelessnetworks, where all nodes are within a transmission rangeof each other. Note, however, that the PRB access method,

0018-9545/$25.00 © 2008 IEEE

GUANG et al.: ENHANCING IEEE 802.11 RANDOM BACKOFF IN SELFISH ENVIRONMENTS 1807

as a misbehavior prevention scheme, can also be applied tomultihop ad hoc networks.1 In addition, in this paper, we donot consider the case of colluding hosts, and we assume, for ourmodel, a saturated network, where each host always has trafficto send. In the rest of this paper, we overview the MAC layerselfish misbehavior in wireless networks in Section II. Then,in Section III, we describe the IEEE 802.11 BEB algorithmand propose our PRB algorithm. Section III-C illustrates themethod for facilitating the detection of selfish behavior withthe cooperation of PRB. In Section IV, we present our PRBanalytical model based on 3-D Markov chains. We considerboth the normal and attack cases for BEB and PRB models.Moreover, we validate our analytical model by implement-ing PRB using NS-2 in Section V. Section VI presents thesimulation results using NS-2. Finally, Section VII concludesthe paper.

II. SELFISH BEHAVIOR IN THE MAC LAYER

The IEEE 802.11 distributed coordination function (DCF)mode [1] combines carrier sense multiple access/collisionavoidance with a request-to-send/clear-to-send (RTS/CTS)handshake to avoid collisions.

With the implementation of the MAC protocol in softwarerather than hardware or firmware in network access cards, itis easy to modify the protocol by a selfish or greedy node[7], [10], [21] to increase their own share of the commontransmission resource. A selfish node may adjust or manipulateits backoff mechanism in different ways to access the channelwith a higher probability. One way is to choose a small backoffvalue rather than a valid generated random number by thebackoff algorithm, e.g., using the range [0, (CW/2)] rather than[0, CW ] (CW is the size of contention windows) or alwaysgenerating a small random value regardless of the range. Inthe presence of a collision or busy medium or retry, the selfishnode will have more chances of winning the channel than theother nodes. A selfish node may also set a longer time durationthan the actual transmission time in its RTS/CTS. Those nodesthat overhear the exchange will have to accordingly adjust theirnetwork allocation vector and, consequently, defer for a longertime before transmission. Furthermore, they can even adjust thedistributed interframe space or the short interframe space timeto further exacerbate the unfairness.

Kyasanur and Vaidya [18] showed that such selfish mis-behavior can seriously degrade the performance of the net-work, and accordingly, they proposed some modifications forthe protocol (e.g., by allowing the trusted receiver to assignbackoff values rather than the sender) to detect and penalizemisbehaving nodes. Similarly, Raya et al. [21] addressed thesame problem and proposed a system, called DOMINO, todetect greedy misbehaviors in IEEE 802.11. However, thesesystems can easily be exploited by new and smart selfish nodes,as presented in [10]. Furthermore, most of these systems do notapply for MAC misbehaviors in MANETs.

1In multihop ad hoc networks, there are serious problems arising from ex-posed and hidden terminals that require the design of a more robust misbehaviordetection system. This is outside the scope of this paper.

III. PROPOSED SCHEME

Our proposed algorithm is designed to require minimalmodifications to the IEEE 802.11 DCF mode and mitigatethe negative impacts in the presence of misbehaved nodesthat manipulate the selection of cw to achieve selfish goals.In this section, we present our scheme PRB that is based onthe modifications of BEB. Finally, we explain how PRB canbe incorporated into IEEE 802.11 and the existing detectionsystems to mitigate and detect selfish MAC layer misbehaviors.

Algorithm 1 PRB1: CW 0

lb = CW deflb = 1

2: cw0 ← [0, 2imin − 1]3: for each sending packet Pi do4: if αl × cwi < Wt then5: if cwi == 0 then6: CW i+1

lb = CW speclb

7: else8: CW i+1

lb = αl × CW ilb

9: end if10: else11: CW i+1

lb = CW deflb = 1

12: end if13: cwi+1 ← [CW i+1

lb − 1,min(2imin+nf , 2imax)−1]14: end for

A. Preliminary

The key idea of PRB is to adjust, in a predictable manner,the lower bound of the contention window (denoted as CWlb;initially, CWlb = 0) to enhance the per-station fairness inselfish environments, i.e., to ensure that well-behaved stationsreceive a fair share of the channel bandwidth. Moreover, theproposed protocol will ensure that selfish nodes will (and unlikeBEB) easily be detected and punished if they deliberatelydo not follow the operation of PRB. We found that the pre-dictable change of the lower bound can result in a significantperformance improvement of well-behaved stations (in termsof throughput, delays, and fairness) in the presence of selfishstations. Moreover, the method can also improve the overall net-work throughput in congested nonselfish environments whilemaintaining a good fairness index.

Other predictable backoff schemes, e.g., IdleSense [13] andlinear/multiplicative increase and linear decrease (LMILD)[8], have been proposed as well. LMILD is only designedto improve the network throughput (i.e., does not deal withfairness) in a congested environment. Its contention range canbe predicted as follows. In case of a failed transmission, CWub

is either linearly increased or multiplicatively increased witha factor of 1.5 instead of 2. Upon a successful transmission,instead of directly resetting CWub, CWub will be linearlyreduced, whereas CWlb is always set to 0. IdleSense [13], onthe other hand, is a novel access method derived from the IEEE802.11 DCF in which all stations use similar value for thecontention window to benefit from good short-term fairness.The analysis of IdleSense showed a high throughput, a lowcollision overhead, a low delay, and a better channel access


fairness. However, unlike IdleSense, our PRB protocol focuseson changing the contention window lower bound to achievehigher throughput and better channel access fairness for well-behaved nodes in the presence of selfish hosts. Furthermore,PRB will prevent selfish hosts from achieving a higher advan-tage of the channel over well-behaved hosts by manipulatingaccess protocol parameters.

B. Principle of PRB

PRB operates as follows. Initially, a node with a data packetto transmit randomly chooses a cwi from [0, CWmin]. Upon asuccessful data transmission, if both cwi and αl × cwi are lessthan Wt,2 a lower bound of the contention window for the nextcwi+1 selection will be assigned as CW i+1

lb = αl × CW ilb. In

case cwi is selected as 0, CWlb is set to a prespecified valueCW spec

lb . Otherwise, CWlb will be set to a default value.3

Therefore, the node needs to select cwi+1 from [CW i+1lb − 1,

CWmin] for the next transmission. In the presence of afailed transmission due to collision or packet errors, theupper bound CWub is doubled, and cw is selected from[CW i+1

lb − 1,min(2imin+nf − 1, 2imax − 1)], where nf is thenumber of failed transmissions. Finally, CWub continues toincrease until it reaches CWmax.

C. Detection and Reaction Based on PRB

Contrary to a previous paper [18], the role of a receiver is toonly detect a selfish sender to eliminate the exploits introducedby untrusted partners [10].

First, we consider a selfish node that only aims at obtain-ing higher throughput by choosing smaller cw. An Rx4 cancompute Bi

act, which is the actual backoff time duration, foreach received frame by using the methods described in [18] and[21]. The condition CW i

act < CW ilb directly identifies a Tx as

misbehaved, where CW ilb is the lower bound. Unlike [18] and

[21], there is no need for λ5 to ensure correct detection becausethe value of CW i

lb is deterministic and can easily be calculatedby the Rx through monitoring the transmissions of Tx. Conse-quently, the detection procedure is simple and accurate. In thepresence of interference, we can improve the positive alarmsby introducing multiple-frame monitoring in addition to per-frame monitoring. Note that choosing CW i

act ≥ CW ilb is not

sufficient to indicate a well-behaved Tx. However, even if aselfish Tx tries to keep on selecting a small cw by applying thestrategies that we described in [10], as long as it follows PRB(i.e., choosing cwi larger than CW i

lb), the negative impacts aresignificantly mitigated, as will be shown in Section V. Thesame detection methods can be applied to detect a selfish nodethat selects a larger cw to reduce power consumption. This

2αl is a multiplicative factor used to increase the lower bound of CW(αl >1). Wt is the threshold to reset the current lower bound of CW to 0.

3Note that the default value of CWlb is set to CWdeflb = 1.

4In a single handshaking process, a node will play two different roles, i.e.,transmitter (Tx) or receiver (Rx). To avoid confusion in the following sections,we use transmitter (Tx) to refer to the source of a MAC data frame, whereasreceiver (Rx) refers to the destination of a MAC data frame.

5λ is a configurable parameter that is selected according to the desired correctdiagnosis ratio and misdiagnosis ratio.

requires modifications to the PRB by introducing an upperbound CWub. Accordingly, a node that keeps on choosinglarger cw will be identified if CW i

act > CW iub. Moreover, as

the detection is based on per-frame monitoring, it is faster thanother detection systems [18], [21] (where the detection systemneeds to be triggered after statistical analysis for a monitoringperiod, which includes more than one frame transmission). Itis to be noted that in a multihop network, problems arisingfrom hidden and exposed terminals cause impairments for thedetection system. For instance, when a transmitter is trying tocommunicate with a receiver that is exposed to another com-munication (of which the transmitter is not aware), the receivermay falsely designate the transmitter as a selfish host becausethe latter did not back off for a longer period. Accordingly, wesuggest that a better detection system must be designed to dealwith problems arising from selfish, hidden, and exposed hosts.It is important, however, to mention that the PRB access methodis a prevention scheme. The design for a detection system in amultihop network is a challenging research that is outside thescope of this paper.

IV. MODELING AND ANALYSIS

In this section, we present the details of our proposedthroughput model for BEB and PRB protocols.6 We haveused the same assumptions made in [4] in our model. First,we present the analysis of BEB. Second, we develop themathematical model for PRB. Both normal and attack casesare considered in the models. We first study τ , which is thetransmission probability of a single station during a randomlyselected timeslot. Then, we express the throughput model forthe whole network as a function of the variable τ .

A. Normal Case

1) Analysis of BEB: We briefly present the model developedby Bianchi for BEB [4]. Let b(x) be the stochastic processthat represents the backoff time counter for a given stationand s(x) be the stochastic process that represents the backoffstage. Let Wi be the contention window size, which is definedas Wi = 2iW0, where W0 = CWmin is the initial contentionwindow value, and let Wm = CWmax

7 be the maximum valueof the contention window. Denote p as the conditional collisionprobability, which is constant and independent regardless of thenumber of retransmissions incurred. Therefore, the system canbe modeled as a 2-D stochastic process s(x), b(x).

τ is given by

τ =2(1 − 2p)

(1 − 2p)(W0 + 1) + pW0 [1 − (2p)m]. (1)

2) Analysis of PRB: Recall that in PRB, a station adjustsits lower bound (CWlb) for the contention window for thenext stage upon each successful transmission if the chosencwi at stage i is lower than the specified Wt, as explained in

6Initial work has been presented in [11].7For convenience, we interchangeably use CW and W to refer to the

contention window size adopted by a station.


Algorithm 1. Therefore, we let l(x) be the stochastic processthat represents the lower bound for a particular contentionstage.

First, we have Wt ∈ [0,W0 − 1], where Wt is the threshold,and W0 is the minimum CW upper bound.8 When t = 0,PRB operates the same way as BEB. Let j be a variablethat represents the lower bound stage of a contention windowgiven by

2jmin − 1 = 0 =⇒ jmin = 0 (2)

αl · 2jmax =Wt =⇒ jmax =⌈log2

Wt

αl

⌉. (3)

We define Mj as the lower bound of a particular contentionwindow at stage j by

Mj = 2j (4)

where j ∈ [0, log2(Wt/αl)]. Let Mt be the maximum lowerbound stage, and define it as follows:

Mt = 2t = Wt/αl. (5)

With these, we obtain a 3-D stochastic process l(x),s(x), b(x) to model the protocol behavior as a discrete-timeMarkov chain depicted in Fig. 1. In the Markov chain, the onlynon-null, one-step transition probabilities are given by

Pj, i, k|j, i, k + 1 = 1Pj + 1, 0, k|j, i, 0 = 1−p

W0−Mj+1+1 × Mt−Mj+1Wi−Mj+1

P0, 0, k|j, i, 0 = 1−pW0

× Wi−Mt

Wi−Mj+1

Pj, i, k|j, i − 1, 0 = pWi−Mj+1

Pj,m, k|j,m, 0 = pWm−Mj+1

(6)

where the corresponding ranges for j, i, k are given by

j ∈ [0, t], i ∈ [0,m], k ∈ [0,Wi − 2]j ∈ [0, t − 1], i ∈ [0,m], k ∈ [Mj+1 − 1,W0]j ∈ [0, t], i ∈ [0,m], k ∈ [0,W0 − 1]j ∈ [0, t], i ∈ [1,m], k ∈ [Mj − 1,Wi − 1]j ∈ [0, t], i ∈ [m,m], k ∈ [Mj − 1,Wm − 1].

(7)

The transition probabilities in (6) are explained as follows.• P(j,i,k+1)→(j,i,k): A station will decrease its backoff

counter by 1 after it senses that the channel is idle for eachtimeslot; otherwise, it will freeze its counter.

• P(j,i,0)→(j+1,0,k): If a station succeeds in the current trans-mission, it will increase its current CW lower bound fromj to j + 1 only if the current selected cw is less than Mt.

• P(j,i,0)→(0,0,k): Once the backoff counter k reaches 0, astation can start transmitting. Upon the successful trans-mission, if the current selected cw is greater than or equalto Mt, the CW lower bound will reset to 0.

• P(j,i−1,0)→(j,i,k): If a station fails to transmit a packet(due to collision or packet error), it will increase thecurrent CW upper bound while keeping the lower boundunchanged.

8In the following discussion, without loss of generality, we always assumethat Wt ≤ W0. The case where Wt > W0 will not be discussed in this paper.

• P(j,m,0)→(j,m,k): If a station faces failed transmission andits current CW upper bound reaches the maximum valueCWmax (i = m), it will keep on using the same upperbound CWmax until successful transmission.

Let bj,i,k =limx→∞Pl(x)=j, s(x)= i, b(x)=k, j∈ [1, t],i ∈ [0,m], k ∈ [Mj − 1,Wi − 1] be the stationary distributionof the chain. We consider the following two different cases inour study to derive the station transmission probabilities: 1) thecase when j = 0 and 2) the case when j > 0.

a) Case when j = 0: In the presence of a failed trans-mission, a station will double its upper bound until it reachesCWmax. According to the state balance condition, we can thenwrite (the derivation is shown in the Appendix)

b0,i,k =Wi − k

Wi

t∑j=0

m∑i=0

bj,i,0(1−p)(Wi−Mt)

Wi−Mj+1 , i = 0

p · b0,i−1,0, 0 < i < mp · (b0,m−1,0 + b0,m,0), i = m

(8)

where k ∈ [0,Wi − 1].b) Case when t ≥ j > 0: bj,i,k is given by the following

(the derivation is shown in the Appendix).• If k ≥ Mj − 1, then

bj,i,k =Wi−k

Wi−Mj +1

m∑i=0

bj−1,i,0Mt(1−p)Wi−Mj+1 , i=0

p · bj,i−1,0, 0<i<kp · (b0,m−1,0+b0,m,0), i=m.

(9)

• If 0 ≤ k < Mj − 1, then

bj,i,k = bj,i,k+1. (10)

The ratio (Wi − k)/(Wi − Mj + 1) accounts for the dis-tribution of probabilities for each state in the correspondingstages j and i. As shown in Fig. 1, when the stage movesfrom the left side to the right side, the probability decreasesby [1/(Wi − Mj + 1)] (k ≥ Mj − 1). However, unlike BEB,when 0 ≤ k < Mj − 1, bj,i,k always gets a single input frombj,i,Mj−1, which means that PRB eliminates the chance that thestate bj,0,k (k < Mj − 1) is directly selected from bj−1,i,0.

Clearly, it is not trivial to obtain a closed-form expressionfor bj,i,k. Hence, we solve the system throughput model via nu-merical simulations based on Matlab. In Section V-A, we givedetails about the procedure of the numerical simulations. From(9) and (10), we can express bj,i,k into a function of two vari-ables b0,0,0 and p. From the normalization condition (11), wecan write 1 = b0,0,0 · Φ(p), where Φ(p) is a function of p only.

Therefore, b0,0,0 is equal to [1/Φ(p)], which can bederived from

1 =t∑

j=0

m∑i=0

Wi−1∑k=0

bj,i,k

=m∑

i=0

Wi−1∑k=0

b0,i,k +m∑

i=0

Mj−2∑k=0

bj,i,k +m∑

i=0

Wi−1∑k=Mj−1

bj,i,k

=t∑

j=0

m∑i=0

Wi − Mj + 22

bj,i,0. (11)


Fig. 1. Markov chain model (3-D). PRB.

Single Station Transmission Probability: Now, we can ex-press the probability τ that a station transmits in a randomlyselected timeslot. Once the backoff time counter reaches 0 (i.e.,k = 0), a transmission will start regardless of the backoff stage(∀i) and lower bound (∀j), i.e.,

τ =t∑

j=0

m∑i=0

bj,i,0 =1

1 − p

t∑j=0

bj,0,0 == Ψ(p,W0,m, t, b0,0,0).

(12)

From (8), (12), and (A-15), we can see that τ is a functionof only one unknown variable p. The other variables (W0,m, t)have known values, and b0,0,0 can be numerically solved, asexplained in Section V-A.

Now, we assume that n identical stations independentlyshare the channel. Therefore, the probability that each stationencounters contentions is independent of the other stations.Moreover, the probability τ that a station transmits a packet

during a timeslot is the same for the n stations. Then, p isthe probability that at least one station transmits in the sametimeslot (i.e., collision), which is given by

p = 1 − (1 − τ)n−1. (13)

Finally, the functions (12) and (13) represent a nonlinearsystem with two unknown variables τ and p. It can be provedthat this system has a unique solution.

• The expression of p in (13) is continuous and is monotone,increasing with τ , which is proved by the fact that thefirst-order derivation of p is always larger than 0: p(1) =(n − 1)(1 − τ)n−2 ≥ 0. Moreover, τ = 0 when p = 0,and τ = 1 when p = 1.

• The expression of τ in (12) is continuous with p. To obtainthe unique solution, τ needs to be continuously decreasingwith p, i.e., more contentions lead to a lower transmission


probability. Note that, from (12) and (A-15), τ reachesthe maximum value when p → 0 and the minimum valuewhen p → 1. Furthermore, from our numerical results, italso shows that τ is continuously decreased when p isincreased.

Total Network Throughput Model: Here, we present thetotal network throughput model for both BEB and PRB. Somenotations are given as follows:

• Ptr: the probability that there is at least one transmissionin a considered timeslot, which is given by

Ptr = 1 − (1 − τ)n (14)

• Ps: the probability that a transmission is successful by theprobability that exactly one station transmits on the chan-nel, based on the fact that at least one station transmits,which is given by

Ps =nτ(1 − τ)n−1

Ptr=

nτ(1 − τ)n−1

1 − (1 − τ)n(15)

• E[PL]: the average packet payload size;• To: the duration of an empty timeslot;• Ts: the average time that the channel is sensed busy for a

successful transmission;• Tf

9: the average time that the channel is sensed busy for afailed transmission;

• Γ: the normalized system throughput.Hence, the expression for the system throughput Γ is given

by (16). Note that the difference between BEB and PRB is thetransmission probability τ , which is defined by (1) and (12),respectively, as

Γ =PsPtrE[PL]

(1 − Ptr)To + PtrPsTs + Ptr(1 − Ps)Tf. (16)

Individual Node Throughput Model: Here, we present theindividual node throughput model for both BEB and PRB.Some notations are given as follows:

• Usv : the probability that a station v successfully transmits

during a timeslot, which is given by

Usv = τv

n∏q=1,q =v

(1 − τq) (17)

• Us: the sum of the probabilities of successful transmis-sions for all stations, which is given by

Us =n∑

v=1

Usv (18)

• Uo: the probability that the channel is idle, which isgiven by

Uo =n∏

v=1

(1 − τv) (19)

• Uf : the probability that collisions occur, which is given by

Uf = 1 − Uo − Us. (20)

9Refer to [4] for the mathematical expression of E[PL], To, Ts, and Tf .

Hence, the throughput Ω(v) of an individual station v isgiven by

Ω(v) =Us

vE[PL]UoTo + UsTs + UfTf

. (21)

B. Attack Case

In Section IV-A, the presented models for BEB and PRBare under the assumption of a saturated channel. Because theobjective of a selfish station is to maximize its own throughput,it will always tend to use the full channel capacity (i.e., thesystem will operate at the saturation point).1) Analysis of BEB: From (1), we can model several selfish

strategies that manipulate CW . For example, a selfish nodechooses cw from [0, γCWmin] (0 ≤ γ < 1),10 instead of from[0, CWmin], and can be modeled as

τBEBs=

2(1 − 2ps)(1 − 2ps) (W γ

0 + 1) + pW γ0 [1 − (2ps)m]

(22)

where W γ0 represents γWmin, and ps is the collision probability

for a selfish node. Another example is that a selfish node selectsCWmin = CWmax = Ws, where Ws is the contention windowspecified by the selfish node. The transmission probability of aselfish node is given by

τBEBs=

2(1 − 2ps)(1 − 2ps)(Ws + 1) + psWs[1 − (2ps)i]

. (23)

In the case where a selfish node always assigns CWmax toCWmin, which can be specified by Wβ , there will be no backoff(i = 0), and (23) can be simplified as

τBEBs=

2Wβ + 1

. (24)

From (22) and (23), it is clear that a selfish node will havemore chances of accessing the channel than in the normal case.Furthermore, the access probability for a well-behaved node isgiven by

τBEBw=

2(1 − 2pw)(1 − 2pw)(W0 + 1) + pwW0 [1 − (2pw)i]

(25)

where pw is decided by

pw = 1 − (1 − τw)n−ns−1(1 − τw)ns (26)

where ns is the number of selfish stations in the network.2) Analysis of PRB: Similar to BEB, we can model selfish

attack strategies by manipulating (12), (A-3), and (A-15). Forexample, a selfish node chooses cw from [0, γCWmin] (0 ≤γ < 1) instead of [0, CWmin] and can be modeled as

b0,0,0 =t∑

j=0

m∑i=0

bj,i,0 ·(1 − p) (W γ

0 − Mt)W γ

0 − Mj + 1(27)

10γ is the misbehavior coefficient to determine the percentage of selfishmisbehavior.


TABLE IPARAMETERS FOR ANALYTICAL AND SIMULATION RESULTS

and

bj,0,0 =m∑

i=0

bj−1,0,0(1 − p)(Mt − Mj−1 + 1)pi

W γ0 − Mj−1 + 1

. (28)

Equations (27) and (28) account for the fact that a selfishstation selects a Ws larger than Mt but smaller than the standardvalue CWmin. If a selfish station selects a Ws smaller than Mt,it has to increase its lower bound upon a successful transmis-sion with probability 1 instead of [(1 − p)(Mt − Mj−1 + 1)]/(2iWs − Mj−1 + 1), which is always less than or equal to 1.Clearly, in PRB, a selfish station selecting a smaller cw willexperience more states to transit back to state b0,0,0, whereas inBEB, a selfish station can immediately transit to b0,0,0 upon asuccessful transmission.

V. PERFORMANCE ANALYSIS

In this section, we first validate our proposed PRB modelin Section V-A. In Section V-B, we compare the performancebetween PRB and BEB.11

A. Model Validation

We validate our proposed 3-D Markov chain throughputmodel for PRB by comparing the numerical results (Matlab)with the simulation results using the network simulatorNS-2 [9].

We compare the numerical and simulation results under boththe basic and the four-way handshaking mechanisms. For thispurpose, we consider an ad hoc network with ns = 1 selfishstation out of n stations, where n stations are randomly gener-ated in a 100 m × 100 m area, and each station is within thetransmission range of the others. The simulated stations have abuffer size of 64 packets. All the parameters that are used forperformance evaluations are summarized in Table I. In Fig. 2,we present the total network throughput of BEB (designated as“BEB”) and PRB (designated as “PRB”) under the normal case,i.e., no selfish attack. The model results (designated as “Mod”)are quite close to the simulation results (designated as “Sim”).Clearly, in the two figures, we can see that the performance

11All the evaluations are under the saturated conditions. The development ofmodels under nonsaturated conditions will be left for future work.

of PRB is comparable to that of BEB and performs evenslightly better than BEB under both basic access and RTS/CTShandshaking scheme. This is due to the fact that a well-behavedstation will not continuously choose a smaller cw, and hence, afew time increments of CW will not cause performance degra-dation. Moreover, as a station under PRB cannot always capturethe channel by deliberately choosing a smaller cw, this allowsother stations to have more chances of accessing the channel,thus increasing the fairness between stations, as well as improv-ing the throughput by reducing the collision probabilities.

B. Efficiency of PRB

In this section, we evaluate the efficiency of our proposedalgorithm PRB in mitigating the negative effects that the manip-ulation of cw could cause on well-behaved hosts. We comparethe performance of BEB and PRB using the following twodifferent scenarios: 1) normal case and 2) attack case.12

1) Attack Case: Without loss of generality, we consider aselfish station that tries to capture the channel by selectinga smaller cw from [0, γCWmin] (0 ≤ γ < 1). In PRB, as theselfish station intends to avoid quick and easy detection, as ex-plained in Section III-C, it has to increase the lower bound if itselects a cw less than the threshold upon a successful transmis-sion. We consider the worst case for PRB, i.e., a selfish stationwill always choose a cw that is equal to the current lower bound.

Fig. 2 compares the throughput obtained by a selfish flow(SF) (designated as “Selfish”) and a well-behaved flow (WF)(designated as “Normal”) using our proposed PRB with thatobtained using the IEEE BEB when varying the misbehaviorcoefficient γ, i.e., γ is equal to 0.2 and 0.6, respectively.For convenience, the figures also show the total networkthroughput when using both schemes. As seen in Fig. 2(a),when γ = 0.2, corresponding to Ws = 6, the normalizedthroughput for an SF has dramatically decreased by 60%(from 53% using BEB to 24% using PRB) when the numberof stations is five. As the number of stations increases, PRBstill performs better than BEB. Similar results can be obtainedfrom Fig. 2(b). When the misbehavior is not very severe, i.e.,γ = 0.6 (which corresponds to CWmin = 19) for a selfishstation, the throughput of an SF is slightly lower than anormal flow, indicating that the fairness index13 is close to 1.Moreover, in terms of the total network throughput (designatedas “Total”), PRB slightly outperforms BEB by 4% (γ = 0.2)and 6% (γ = 0.6). Note that, when BEB is used, the SFachieves much higher normalized throughput over the normalflow of a well-behaved station [as can be seen in Fig. 2(a)].

Fig. 3 shows the throughput analysis for PRB when varyingthe threshold Wt and the increment factor αl (αl > 1). Asshown in Fig. 3(a), the throughput for an SF continuouslydecreases as Wt increases. This is because a selfish stationneeds to experience more states before it resets its contention

12For brevity, we only consider the basic access case in this section, i.e., thetwo-way handshaking data/acknowledgment.

13Jain’s fairness index [17] is given by FJ =[(∑N

i=1Ti)

2)/

(N∑N

i=1T 2

i

)], where N is the number of flows, and Ti is the throughput of

a flow i. FJ is equal to 1 when all the flows equally share the channel capacityand is equal to 1/N when a single flow occupies the full bandwidth.


Fig. 2. Selfish/normal flow throughput versus number of stations (attack case, Wt = 32) when varying γ. (a) γ = 0.2. (b) γ = 0.6.

Fig. 3. Efficiency of PRB parameters: Wt and αl (γ = 0.2). (a) Throughput analysis when varying Wt. (b) Throughput analysis when varying αl.

window to [0, CWmin] with a higher Wt. For example, whenWt = 4, the throughput for an SF is 0.46, whereas the normal-ized throughput is 0.24 when Wt = 32. This is due to the factthat a selfish node can still obtain more channel bandwidth byselecting a value that is equal to the threshold if Wt is chosensmall (a small Wt allows a fast resetting of the lower bound ofthe contention window back to 0). Therefore, Wt is required tobe selected to a relatively larger value. Similarly, as shown inFig. 3(b), the throughput for an SF continuously decreases asαl decreases. For example, when αl = 8, the throughput for anSF is 0.46, whereas it is 0.24 when αl = 2. This is due to thefact that, given a fixed threshold Wt, a smaller αl requires anode to experience more lower bound stages than a larger αl,which may result in a lower bound reaching the threshold (Wt)very fast.

VI. NS-2 SIMULATION AND ANALYSIS

We use NS-2 [9] to evaluate the performance of our proposedalgorithm PRB. Although PRB is efficient regardless of the

attack strategies that an attacker might apply, e.g., deliberatelychoosing a smaller or larger cw, in the following discussion, weonly consider a selfish node that chooses a smaller cw to im-prove its own throughput while deteriorating the performanceof other WFs.

A. Simulation Setup

Simulation Scenario: Consider a ring network with one nodeat the center and the rest of the nodes uniformly distributed onthe circle. The ring radius is 200 m. All the nodes are fixed.The transmission range for each node is 250 m, and the carriersense range is 550 m. Each node on the circle is the source ofa single data flow, and the source of each flow has the centerpoint as the destination. The traffic type is constant bit rate(CBR). The packet size is 512 B/packet, and each flow hasthe same packet arrival rate. The channel bit rate is 2 Mb/s.The total time for each simulation run is 500 s. Each datapoint on the following figures is averaged over five runs. The


Fig. 4. BEB versus PRB. Load tests—Overall network performance. (a) Total network throughput versus offered load. (b) Packet delivery ratio versus offeredload. (c) Average packet delay versus offered load. (d) Fairness index versus offered load.

default parameters for PRB are configured as follows: αl = 2;Wt = 31; and CW spec

lb = 4.To model the selfish misbehavior, we consider the worst

case of PRB, i.e., a selfish source node always chooses abackoff cw from [CWlb + 0, CWlb + γCWmin]14 (γ ≥ 0),where γ is the misbehavior coefficient, indicating the percent-age of deviation from the standard with a default value 0.1.Simulation Metrics: We use the following metrics to study

the performance of our proposed approach:

1) Packet delivery ratio: the ratio of the data packets that aresuccessfully delivered to the destination by all the flowsto those generated by the sources;

2) Average packet delay: the average end_to_end delay foreach successfully delivered data packet, which includes

14PRB is also efficient in mitigating other smart cw selection strategies, aswe explained in Section III-B. For brevity, only the naive attack strategy [10]wherein a node always selects a smaller cw, i.e., 0 ≤ γ ≤ 1, is discussed inthis section.

all the possible delays caused by route buffering, MACinterface queue, and retransmission delays;

3) Fairness index: the Jain’s fairness index, as definedin [17].

B. Simulation Results

1) Manipulation of CW: We start our evaluation by compar-ing the overall network performance between BEB and PRBwhen varying the offered load. First, we observe that the totalnetwork throughput for BEB and PRB are similar under thenormal case (designated as “Normal”) and the attack case(“Attack”), as shown in Fig. 4(a). Initially, when the channelis less congested, BEB and PRB identically perform in bothcases. However, as the channel becomes more congested, e.g.,when the offered load is 16 386 kb/s, BEB and PRB exhibitdifferent behaviors. In either of the two cases (normal or attack),we notice that an increment of the CW lower bound (as inPRB), instead of resetting to 0 (as in BEB), will not cause


Fig. 5. BEB versus PRB. Load tests—Per-flow performance. (a) Per-flow throughput versus offered load. (b) Per-flow end-to-end delay versus offered load.

performance penalty. On the contrary, it will slightly improvethe total network throughput, e.g., the performance marginbetween BEB and PRB ranges from 20 to 50 kb/s. This isdue to the fact that, in a congested environment, a node maysuccessfully send a packet after the ith15 contention stage.Directly resetting the lower bound to 0 might cause this node to,again, experience the same contention procedures if the currentchannel maintains the same congestion conditions. Therefore, agentle decrement in CW is preferred [3], [8]. Here, PRB sharesa similar property with LMIMD [8] and SD [3] by forcing anode to choose a relatively large cw than BEB. However, unlikeLMIMD and SD, which both focus on adaptively changing theupper bound of CW , PRB uses a different method by adjustingthe lower bound. With proper selection of PRB parameters,we can achieve a much better throughput gain, as will bediscussed later. Similar results can be seen in Fig. 4(b), whichshows the packet delivery ratio when varying the traffic load.Clearly, when the channel is less congested, in either BEB orPRB, all the generated packets can be successfully deliveredto the destination, which yields a 100% packet delivery ratio.Alternatively, the delivery ratio decreases as the offered loadincreases.

Fig. 4(d) shows the results of fairness when varying theoffered load. It is clear that BEB and PRB performs similarlywhen the offered load is below 2000 kb/s with or without attack.This first proves the fact that selfish behavior has little impact ina less congested environment. Next, the fairness index of BEBsharply decreases as the offered load increases in the presenceof attacks, whereas PRB achieves higher and stable fairnessindex at higher loads, e.g., FBEB

J = 0.078 and FPRBJ = 0.734

when the load is 16 386 kb/s. Once a node starts acting selfishly,it will increase its chance of capturing the channel by selectinga much smaller cw, particularly in a congested environment.This results in a continuous backoff of its neighboring nodes,which will allow the SF to dominate the medium. Hence, the

15See to Section IV for details.

throughput of an SF will keep increasing as the load increases,which leads to a severe throughput degradation of WFs, partic-ularly when the channel reaches to a saturated condition. Fig. 5shows these results.

Fig. 5(a) shows the performance comparison between oneSF and two WFs (WF: WF1 and WF2). Clearly, in BEB, theSF throughput continuously increases as the load increases,whereas the throughput for WF1 and WF2 is close to 0. This,again, shows that an SF can successfully exploit the protocoland take unfair advantage over other well-behaved hosts. How-ever, in PRB, an SF throughput is significantly reduced and keptstable as the load changes, e.g., the SF throughput drops from743.9 (BEB) to 167.6 kb/s (PRB) when the load is 16 386 kb/s.In addition, in PRB, the delay for an SF is close to that of aWF and is kept stable at various loads. Moreover, the selfishbehavior also results in a continuous backoff of neighboringnodes, which will significantly reduce packet delays for the SFand increased delays for the WFs, as shown in Fig. 5(b). Forexample, in the presence of attacks, in BEB, the averaged delayof WFs is 12 s when the traffic load is 16 386 kb/s, whereas inPRB, the delay is reduced to 5 s. Because the SF occupies alarge portion of the total network bandwidth, e.g., 80% whenthe load is 16 386 kb/s, it will also lead to a decreased averagedelay for the whole network, as shown in Fig. 4(c). Hence, PRBis fairly efficient in dealing with selfish hosts in a congestedenvironment.

Fig. 6 shows the instantaneous throughput of a 20-flowscenario, where each flow is transmitting at 100 packets persecond (pkt/s). As shown in Fig. 6(a), under the normal case, wesee that there is no difference between BEB and PRB in terms ofthe total network throughput and the per-flow throughput. In thepresence of a selfish host, the total throughput of BEB becomesunstable, e.g., the instantaneous throughput drops from 1000to 269 kb/s during the period between 60 and 70, as shownin Fig. 6(b). This is because, when a selfish node selects asmaller CWmin, the selection of a cw in the presence of a failedtransmission will be bounded by a smaller CW upper bound(2iCWmin). This results in a selfish host continuously selecting


Fig. 6. Instantaneous throughput. BEB versus PRB (normal versus attack). (a) Normal case. (b) Attack case.

a very small cw in a congested environment and may frequentlycollide with the other stations for a short period. However, inPRB, the choice of a new cw is bounded by both the lower andupper bounds. Therefore, a selfish host cannot keep on selectinga sequence of small cws while avoiding the detection system.Fig. 6(b) shows that the total throughput of PRB is stable and iskept at 1000 kb/s. Furthermore, PRB successfully mitigates theimpacts of the selfish host by reducing its throughput by 71%,i.e., from 800 (BEB-SF) to 230 kb/s (PRB-SF).

We next explore the effects of varying PRB core parameters(i.e., αl and Wt). The results are presented in Figs. 7 and 8.In this scenario, 20 flows exist in the network, with each flowsending packets at a rate of 2000 pkt/s (i.e., highly congestedenvironment). Fig. 7 shows that PRB can slightly improve thenetwork throughput when choosing αl larger than or equalto 1, indicating an increased lower bound. When αl equalsto 2, the network throughput reaches its maximum in eitherthe attack or normal case [Fig. 7(a)]. Furthermore, the use ofαl (αl ≥ 1) in PRB results in better fairness than BEB withor without selfish hosts [Fig. 7(d)]. Another observation isthat Fj decreases as αl increases. This is simply because agreedy increment of the lower bound of CW (i.e., a large αl)will always result in a selfish host very quickly exceeding thethreshold (Wt), which forces the host to reset its lower boundto 0 much faster. Hence, we suggest the use of a smaller αl

for effective performance of PRB. Fig. 8 illustrates the resultswhen the threshold Wt varies. It is shown that when PRB isused, the network throughput continuously improves as Wt

increases. First, we note that in a congested environment, anode may experience several contention stages, which mayresult in doubling the CW upper bound (CWub) upon eachfailed transmission. Directly resetting CWub to CWmin aftera successful transmission might force the node to experiencethe same contention as before. This is because the choice of anew cw is bounded by [CWlb, CWub]. Therefore, the marginbetween CWub and CWlb determines the number of choices(i.e., CWub − CWlb + 1) that a node could have in selecting

a new cw for the next transmission. Given the same numberof contending stations, a small margin clearly results in morecollisions. Hence, a large margin is preferred when the numberof contending stations and the traffic load are high. We shouldmention here that schemes such as LMIMD [8] and SD [3]are designed to address this issue. Another important factorthat determines the contention frequency in any access protocolis CWlb. In PRB, a large CWlb forces a node selecting acw (cw < Wt) for the previous transmission to back off fora longer duration for the next transmission. Alternatively, astation selecting a cw (cw ≥ Wt) will have more chances ofcapturing the channel for the next transmission because it willreset the lower bound of the contention window to its minimum.Therefore, increasing Wt will enable the lower bound to reacha larger value. Those nodes with larger lower bounds will deferfor a longer period, and nodes with smaller lower bounds willdefer for a shorter period. This, hence, will result in splittingof nodes into different groups, each selecting its backoff fromdifferent contention ranges. Accordingly, the collision proba-bility will be reduced and a higher throughput can be achieved,as shown in Fig. 8(a). Notice also that a higher Wt will alsoachieve better fairness [Fig. 8(d)]. A small Wt will yield a quickreset for the lower bound to the default value (i.e., 0), whereasa large Wt will force selfish hosts to select larger cws, givingmore chances for well-behaved hosts to access the channel.Fig. 8(d) clearly shows better fairness as Wt increases. Now,although a larger Wt achieves better fairness and a significantimprovement of the total network throughput, this, however,comes at the cost of increased end-to-end delays due to theincrement of the lower bound, as can be seen in Fig. 8(c).

Fig. 9 compares the performance of BEB and PRB whenvarying the misbehavior coefficient γ for different pairs ofPRB parameters. Clearly, PRB always outperforms BEB interms of achieving fairness among competing flows (betterfairness index) and in terms of the per-flow throughput, morespecifically, the throughput for the SF. As can be seen in thefigure, when γ = 1 (no attack), the throughput for BEB-SF and


Fig. 7. Effects of varying αl. (a) Total network throughput versus αl. (b) Packet delivery ratio versus αl. (c) Average delay versus αl. (d) Fairness versus αl.

PRB-SF is 55.99 and 55.30 kb/s, respectively. However, underBEB, the throughput for SF rapidly increases as its misbehaviorgets more intense. For example, when γ = 0.1, SF gets 367 kb/sat the cost of severe degradation for WF throughput. On theother hand, in PRB, SF obtains 130 kb/s (αl = 2 and Wt = 63)when γ = 0.1.

Fig. 10 illustrates the overall network performance betweenBEB and PRB when varying the number of misbehaved nodes[designated as MN% = (NUMSelfishSTA/NUMTotalSTA)] ina 20-flow scenario (γ = 0.1). First, the results obtained inFig. 10(a) show that the throughput of BEB and PRB decreaseswhen increasing MN%. Because the goal of each SF is tomaximize its throughput, a selfish node will select a small CWas much as possible, which will also exacerbate the unfairnessamong stations. When the number of stations selfishly actingincreases, we observe a degradation in the network performance[Fig. 10(a)]. For example, if we consider the worst case whereeach node in the network does not back off (cw = 0), continu-ous collisions will occur, and consequently, no node can trans-mit any packet, which will ultimately cause the whole network

to collapse. Fig. 10(a) reflects this fact, where the misbehaviorcoefficient (γ) is 0.1. The figure shows that as the MN%increases, the total throughput (for BEB and PRB) decreases.Here, because the selection range for cw is quite small, a largernumber of MNs will obviously cause more collisions. PRB,however, reduces the effects by reducing the probability that anode selects a small cw (cw < Wt); and hence, PRB reducesthe collision frequency caused by multiple selfish nodes bydelaying their access to the channel. In addition, this delay willhelp the well-behaved nodes to increase their access probability,i.e., during the backoff of selfish nodes, there will be fewercontending stations and fewer selfish nodes.

An interesting observation can be seen in Fig. 10(b), i.e., ineither BEB or PRB, the fairness index first decreases; however,it starts to increase when the MN% ≥ 10%. As seen fromits definition, FJ represents the throughput margin betweeneach individual flow, e.g., the higher the margin, the lowerthe FJ . Initially, the appearance of multiple selfish nodes willexacerbate the contention situation for well-behaved nodes,resulting in further reduction of their throughput, which, in turn,


Fig. 8. Effects of varying Wt. (a) Total network throughput versus Wt. (b) Packet delivery ratio versus Wt. (c) Average delay versus Wt. (d) Fairness versus Wt.

Fig. 9. Efficiency of PRB when varying γ. (a) Per-flow throughput versus γ. (b) Fairness versus γ.


Fig. 10. Efficiency of PRB when varying MN%. Total network performance. (a) Total network throughput versus MN%. (b) Fairness versus MN%.

Fig. 11. Probability distribution of the selected contention window cw. Manipulation of CW . (a) Normal case. (b) Attack case.

indicates a reduction in FJ . However, as more and more SFjoin the network, the payoff obtained by previously existing SFwill be reduced. Eventually, as the MN% increases, the networkwill converge to a point in which no significant payoff can beobtained by any SF, i.e., the channel is fairly shared by all thestations, which, again, indicates an FJ close to 1. Note that,although, in this case, all the nodes can access the medium withan equal probability, the number of collisions might increase(due to the small range of CW ),16 which will consequently leadto a degradation of the total network throughput. This reflects,by our results (as shown in Fig. 10), that the total networkthroughput continuously decreases as MN%, although FJ firstdecreases and then starts to rise up.

To further support the results shown above, we illustratethe probability distribution Pcw of the selected contention

16It will depend on the number of contending stations and the position ofthese stations.

window cw in Fig. 11 for both the attack and normal cases,using both BEB and PRB access protocols. In each scenario,the simulation time is 200 s, all the parameters are set totheir default values, and we compute the probability of cw foreach packet as a function of cw, which satisfies the relation∑CWmax

cw=0 Pcw = 1. First, as we can see in Fig. 11(a), PRBreduces the probability of the selection of a small cw becauseof the increment of CWlb, e.g., the probability to choose acw < 10 is reduced by 5%. Second, for contention windowswhere most transmission occur, PBEB

cw is close to PPRBcw . These

two observations show that for a very short period (e.g., fewseconds), PRB slightly degrades the total network throughputby reducing PPRB

cw , where cw is small, and for a long period(which is the general case for most applications), PRB willadaptively change PPRB

cw to increase the throughput, e.g., PPRBcw

is greater than PBEBcw when cw varies from 30 to 60. In the at-

tack case, as shown in Fig. 11(b), SF in BEB maintains a higherPcw than that of the WF. However, via the adjustment of CWlb


in PRB, Pcw of SF has dropped by more than 40% for cw ≤ 3.In addition, it is clear that PPRB

cw of SF has been dramaticallyincreased compared with PBEB

cw and is close to the normal casewhen cw varies from 10 to 30 due to the default Wt. Overall,we conclude that PRB, unlike BEB, could significantly improvethe network performance in the presence of selfish nodes byadaptively changing the CWlb. PRB, however, requires someminor modifications to the standard.

VII. CONCLUSION AND FUTURE WORK

Wireless MAC protocols currently deployed in MANETsuse distributed contention resolution mechanisms for sharingthe wireless channel. In such an environment, selfish hoststhat fail to adhere to the MAC protocol may obtain an unfairshare of the channel bandwidth at the expense of well-behavedhosts. In this paper, we have presented a PRB algorithm thatis capable of reducing the impacts of selfish hosts on thenetwork performance, particularly on well-behaved hosts. PRBis based on the minor modifications for BEB and forces eachnode to generate a predictable backoff interval. Hosts thatdo not follow the operation of PRB are easily detected andisolated. We have analytically evaluated PRB using a 3-DMarkov chain model to derive the system throughput. Usingboth simulation and numerical results, we have shown theaccuracy of the developed model. We have also shown thatPRB has similar performance to BEB under the normal caseand has superior performance and a better fairness index inthe presence of selfish hosts. We have presented a detailedanalysis of both access protocols in the presence and absence ofselfish hosts.

In future work, we intend to further optimize the opera-tion of PRB. For instance, throughput improvement could beachieved by more conservatively decreasing the contentionwindow upon successful transmissions. We will also be look-ing at adaptively changing the threshold Wt to reduce theimpacts of smart selfish nodes. We plan to further com-pare the performance of BEB and PRB with and withoutattack and considering mobility, network size, and traffictype. We also intend to look at the detection and reactionsystems.

APPENDIX

a) Case when j = 0: When j = 0, PRB and BEB behavesimilarly. In the presence of a failed transmission, a station willdouble its upper bound until it reaches CWmax. According tothe state balance condition, we can then write

b0,i,Wi−1 = b0,i−1,0 ·p

Wi

b0,i−1,0 ·p

Wi+ b0,i,1 = b0,i,0 · p + b0,i,0

· (1 − p) ·(

Wi − Mt

Wi+

Mt

Wi

)

b0,i−1,0 · p = b0,i,0 (A-1)

where j = 0, i ∈ (0,m), and k ∈ [0,Wi − 1]. When i = m,b0,m,0 can be written as

b0,m−1,0 · p + b0,m,0 · p= b0,m,0 · p + b0,m,0 · (1 − p)

·[W0(Wm − Mt)

W0Wm+

(W0 − M1 + 1)Mt

(W0 − M1 + 1)Wm

]

b0,m−1,0 · p= b0,m,0 · (1 − p) (A-2)

where j = 0, i = m, and k ∈ [0,Wm − 1]. When i = 0 andk = 0, b0,0,0 is given by

b0,0,0 =t∑

j=0

m∑i=0

bj,i,0 ·(1 − p)(Wi − Mt)

Wi − Mj + 1. (A-3)

Therefore, the relation between b0,i,0 and b0,i−1,0 is

b0,i,0 = pib0,0,0

b0,m,0 =pm

1 − pb0,0,0. (A-4)

By using (A-1) to (A-4), we derive the functions for b0,i,k as

b0,i,k =Wi − k

Wi

t∑j=0

m∑i=0

bj,i,0(1−p)(Wi−Mt)

Wi−Mj+1 , i = 0

p · b0,i−1,0, 0 < i < mp · (b0,m−1,0 + b0,m,0), i = m

(A-5)

where k ∈ [0,Wi − 1].b) Case when t ≥ j > 0: First, from Fig. 1, we can derive

the relation between b1,0,W0−1 and b0,0,0 as

b1,0,W0−1 =b0,0,0Mt(1 − p)

W0(W0 − M1 + 1)

[1 −

(p2

)m

1 − p2

+(p

2

)m

].

(A-6)

Next, we have the relation

b1,0,M1−1 = (W0 − M1 + 1) · b1,0,W0−1

b1,0,M1−1 = b1,0,M1−2 = · · · = b1,0,1

b1,0,1 = b1,0,0 · (1 − p + p) = b1,0,0. (A-7)

From (A-6) and (A-7), we can easily write

b1,0,0 =b0,0,0Mt(1 − p)

W0

[1 −

(p2

)m

1 − p2

+(p

2

)m 11 − p

].

(A-8)

In the presence of a failed transmission due to collision orpacket failure, we have

b1,1,W1−1 = b1,0,0 ·p

W1 − M1 + 1b1,1,W1−2 = b1,1,M1−1 + b1,0,0 ·

p

W1 − M1 + 1b1,1,M1−1 = b1,0,0 · p. (A-9)


As we can see in (A-9), a new random backoff counter isreselected from the range [M1 − 1,W1 − 1] (PRB) instead of[0,W1 − 1] (BEB). Therefore, this reduces the chance that astation keeps on selecting a smaller cw. Next, note that whenthe system reaches a state where k = M1 − 1, it will continueto change states with a probability of 1 until k = 1. Hence,we have

b1,1,M1−1 = b1,1,M1−2 = · · · = b1,1,1. (A-10)

Furthermore, we can derive

b1,1,1 = b1,1,0 · p + b1,1,0 ·(1 − p)(W1 − Mt)W0(W1 − M1 + 1)

· W0

= b1,1,0. (A-11)

In the presence of a failed transmission, we obtain

bj,i,0 = bj,i−1,0 · p (A-12)

where 0 < i < m. Similar to (A-4), we write

bj,m,0 =p

1 − p· bj,m−1,0. (A-13)

Next we consider the transitions between bj−1,i,0 and bj,0,k

as follows:

b2,0,W0−1 = b1,0,0(Mt − M1 + 1)(1 − p)

(W0 − M1 + 1)(W0 − M2 + 1)

+ · · · + b1,m,0(Mt − M1 + 1)(1 − p)

(Wm − M1 + 1)(W0 − M2 + 1)

+(

p

W1 − M1 + 1+ · · · + pm

Wm − M1 + 1

).

(A-14)

By using the relations from (A-6) to (A-14), we write

bj,0,0 =m∑

i=0

bj−1,0,0(1 − p)(Mt − Mj−1 + 1)pi

Wi − Mj−1 + 1. (A-15)

When j = t, upon a successful transmission, the state (t, i, 0)will always go back to (0, 0, 0) with probability 1 − p.

Next, according to the Markov chain regularities, foreach k ∈ [0,Wi − 1], bj,i,k can be written as (A-16), wheret ≥ j > 0.

• If k ≥ Mj − 1, then

bj,i,k =Wi − k

Wi − Mj + 1

m∑i=0

bj−1,i,0Mt(1−p)Wi−Mj+1 , i = 0

p · bj,i−1,0, 0 < i < kp · (b0,m−1,0 + b0,m,0), i = m.

(A-16)

• If 0 ≤ k < Mj − 1, then

bj,i,k = bj,i,k+1. (A-17)

REFERENCES

[1] Wireless LAN Media Access Control (MAC) and Physical Layer (PHY)Specifications, IEEE802.11, 1999.

[2] I. Aad, J. P. Hubaux, and E. W. Knightly, “Denial of service resilience inad hoc networks,” in Proc. ACM Mobicom, Sep. 2004, pp. 202–215.

[3] I. Aad, Q. Ni, C. Barakat, and T. Turletti, “Enhancing IEEE 802.11 MACin congested environments,” in Proc. IEEE ASWN, Boston, MA, 2004,pp. 82–91.

[4] G. Bianchi, “Performance analysis of the IEEE 802.11 distributed coordi-nation function,” IEEE J. Sel. Areas Commun., vol. 18, no. 3, pp. 535–547,Mar. 2000.

[5] S. Buchegger and J.-Y. L. Boudec, “Performance analysis of theCONFIDANT protocol (Cooperation of nodes: Fairness in dynamicad-hoc neTworks),” in Proc. ACM Symp. Mobile Adhoc Netw. Comput.(MOBIHOC), Lausanne, Switzerland, Jun. 9–11, 2002.

[6] M. Cagalj, S. Ganeriwal, I. Aad, and J.-P. Hubaux, “On selfish behavior inCSMA/CA networks,” in Proc. INFOCOM, Mar. 2005, pp. 2513–2524.

[7] A. Cárdenas, S. Radosavac, and J. S. Baras, “Detection and preventionof MAC layer misbehavior in ad hoc networks,” in Proc. ACM SASN,Oct. 2004, pp. 17–22.

[8] J. Deng, P. K. Varshney, and Z. J. Haas, “A new backoff algorithmfor the IEEE 802.11 distributed coordination function,” in Proc. CNDS,Jan. 2004.

[9] K. Fall and K. Varadhan, NS Notes and Documentation, 2002. Tech. Rep.Univ. Calif., Berkeley, LBL, USC/ISI. In Xerox PARC.

[10] L. Guang and C. Assi, “Mitigating smart selfish MAC layer misbehaviorin ad hoc networks,” in Proc. IEEE WiMob, Jun. 2006, pp. 116–123.

[11] L. Guang, C. Assi, and A. Benslimane, “Modeling and analysis of pre-dictable random backoff in selfish environments,” in Proc. ACM/IEEEMSWiM, Oct. 2006, pp. 86–90.

[12] V. Gupta, S. Krishnamurthy, and M. Faloutsous, “Denial of service attacksat the MAC layer in wireless ad hoc networks,” in Proc. MILCOM, 2002,pp. 1118–1123.

[13] M. Heusse, F. Rousseau, R. Guillier, and A. Duda, “Idle sense: An optimalaccess method for high throughput and fairness in rate diverse wirelessLANs,” in Proc. ACM SIGCOM, Aug. 2005, pp. 121–132.

[14] Y.-C. Hu, D. B. Johnson, and A. Perrig, “SEAD: Secure efficient distancevector routing for mobile wireless ad hoc networks,” Ad Hoc Netw., vol. 1,no. 1, pp. 175–192, Jul. 2003.

[15] Y.-C. Hu and A. Perrig, “A survey of secure wireless ad hoc routing,”IEEE Security Privacy—Special Issue Making Wireless Work, vol. 2,no. 3, pp. 28–39, May/Jun. 2004.

[16] Y.-C. Hu, A. Perrig, and D. Johnson, “Ariadne: A secure on-demandrouting protocol for ad hoc networks,” in Proc. MobiCom, Sep. 2002,pp. 12–23.

[17] R. Jain, The Art of Computer Systems Performance Analysis. Hoboken,NJ: Wiley, 1991.

[18] P. Kyasanur and N. Vaidya, “Selfish MAC layer misbehavior in wirelessnetworks,” IEEE Trans. Mobile Comput., vol. 4, no. 5, pp. 502–516,Sep. 2005.

[19] P. Papadimitratos and Z. Haas, “Secure routing for mobile ad hocnetworks,” in Proc. CNDS, 2002, pp. 27–31.

[20] P. Papadimitratos and Z. Haas, “Secure link state routing for mobile ad hocnetworks,” in Proc. IEEE Workshop Security Assurance Ad Hoc Netw.,2003, pp. 379–383.

[21] M. Raya, J. P. Hubaux, and I. Aad, “DOMINO: A system to detect greedybehavior in IEEE 802.11 hotspots,” in Proc. ACM MobiSys, Jun. 2004,pp. 84–97.

[22] S. Zhong, L. E. Li, Y. G. Liu, and Y. R. Yang, “On designingincentive-compatible routing and forwarding protocols in wireless ad-hocnetworks,” in Proc. ACM Mobicom, Cologne, Germany, Aug. 2005,pp. 117–131.

[23] Y. Zhou, D. Wu, and S. Nettles, “Analyzing and preventing MAC-layerdenial of service attacks for stock 802.11 systems,” in Proc. WorkshopBWSA, BROADNETS, Oct. 2004.

Lei Guang (S’05) received the B.Eng. degreein automatic control from Nanjing University ofScience and Technology, Nanjing, China, in 2002and the M.Sc. degree in electrical engineeringfrom Concordia University, Montreal, QC, Canada,in 2005.

He is currently a Network Analyst with Voice &Data Systems Inc., Montreal. His research interestsinclude wireless networking, mobile systems, andoperating system and network security.

Mr. Guang is a Microsoft Certified SystemsEngineer/Microsoft Certified Systems Administrator and a Cisco Professional.


Chadi M. Assi (M’03) received the B.Eng. degreefrom the Lebanese University, Beirut, Lebanon, in1997 and the Ph.D. degree from the City Universityof New York in 2003.

He was a Visiting Scientist for one year at theNokia Research Center, Boston, MA, working onquality of service in optical access networks. InAugust 2003, he was an Assistant Professor withthe Concordia Institute for Information SystemsEngineering, Concordia University, Montreal, QC,Canada, where he is currently an Associate Profes-

sor. His current research interests are in the areas of optical networks, wirelessand ad hoc networks, and security.

Dr. Assi received the prestigious Mina Rees Dissertation Award from theCity University of New York in August 2002 for his research on wavelength-division-multiplexing optical networks.

Abderrahim Benslimane (M’99) received the B.S.degree from the University of Nancy, Nancy, France,in 1987, the D.E.A. (M.S.) and Ph.D. degrees fromFranche-Comte University of Besançon, Besançon,France, in 1989 and 1993, respectively, all in com-puter science, and the accreditation (habilitation)for Ph.D. supervision (higher degree research) atthe University of Cergy-Pontoise, Cergy-Pontoise,France, in 2000.

From 1994 to 2001, he was an AssociateProfessor with the University of Technology of

Belfort-Montbeliard, Belfort, France. Since 2001, he has been a Full Profes-sor of computer science and engineering with the Laboratoire Informatiqued’Avignon, Université d’Avignon, Avignon, France, where he is currently theHead of the Computer Networks and Multimedia Applications Team. He isinvolved in many national and international projects. From 2005 to 2006, hewas on sabbatical leave to conduct research at the Department of ComputerScience, Ecole Polytechnique de Montréal, Montreal, QC, Canada. His researchand teaching interests are in mobile ad hoc and wireless sensor networks. Inparticular, he works on security (e.g., intrusion detection, monitoring, etc.),multicast routing, intervehicular communications, quality of service, and MAClayer performance evaluation. He has refereed several international publications(books, journals, and conferences) in all these domains.

Prof. Benslimane has served on the steering and technical program commit-tees of many international conferences.

enhancing ieee 802.11 random backoff in selfish environments

Documents