Incentive Driven LTE Content Distribution inVANETs

Chuchu Wu, Mario GerlaComputer Science Department

University of California, Los AngelesEmail: {wuchuchu, gerla}@cs.ucla.edu

Nicholas MastronardeElectrical Engineering Department

State University of New York at BuffaloEmail: nmastron@buffalo.edu

Abstract—Content downloading from vehicles is becomingincreasingly popular, but it also meets challenges due to thelimited number of road-side units (RSUs) and the expensivecost of 3G/LTE downloads. Cooperative LTE + vehicle-to-vehicle(V2V) downloading schemes can significantly improve the effi-ciency and reduce costs. However, selfish users are motivated tominimize their own LTE downloading costs, so they would ratherdownload content from their peers, i.e., act as free riders. Suchuncooperative user behavior can severely degrade the perfor-mance. In this paper, we propose an incentive driven cluster-basedcontent distribution scheme that encourages cooperation. Vehicleswith common interests form a cluster and take turns to be thecluster head. The cluster head directly downloads data packetsfrom the Internet and V2V shares the content with surroundingvehicles. We show that this scheme cannot result in fair andefficient operation unless it is combined with an appropriatescheduling/management strategy. To this end, we design a server-assisted key management scheme that invites cooperation and alsoensures fairness and efficiency. The proposed scheme is analyzedwith game theoretic methods and is validated via simulation.

I. INTRODUCTION

In the near future, the proliferation of “smart vehicles”is envisaged to become a reality. Smart vehicle on boardprocessors combine the characteristics of both smartphonesand laptops: they are as “mobile” as smart phones, i.e., havingubiquitous access to the Internet via 3G/LTE networks, butmore powerful in computation and energy supply like laptopcomputers. Besides safe-driving services, content downloadingis expected to be widely popular among drivers and passengerson board just like in traditional networks. Since the number ofroad side units is very limited, the time when vehicles can getaccess to Wi-Fi is very short. For most of the time, vehiclesmust rely on 3G/LTE to download Internet content. However,mobile data customers are increasing, and the amount ofdata requested are getting larger and larger, due to increasingdemands for higher quality of video experience. Consequently,the 3G/LTE resources are getting more and more scarce andthus expensive, resulting in longer delays and higher costs ofcontent downloading.

Often in vehicles at the same location moving towardsthe same direction, people onboard are very likely to beinterested in the same video streaming content. One possiblescenario is a certain entertainment mobile application thatprovides limited options of movies to enjoy on the road.As inter-vehicle communication technology becomes available,cooperative content downloading schemes are proposed to

Fig. 1. Cooperative peer-to-peer content downloading in VANETs: vehiclesof common interest (same color in the figure) form a cluster

leverage this advantage and bring peer-to-peer overlay networkto the vehicular environment. An early example is SPAWN [1],a peer-to-peer content delivery mechanism that utilizes paralleldownload across a mesh of co-operating peers. Given the lim-ited number of available RSUs, vehicles with common interestdownload data from the Internet through cellular base stations,instead of RSUs, and share with each other in ad hoc mode,as shown in Fig. 1. Such cooperative downloading schemesignificantly improves the efficiency of content downloadingand can also reduce delays.

Different from traditional peer-to-peer file sharing sys-tems (like Bit-Torrent) where for each peer, downloadingis beneficial and uploading is costly, in vehicular networks,downloading content via 3G/LTE is expensive and inter-vehiclesharing is just about free. So instead of avoiding the sharing offiles with peers, selfish nodes in vehicular networks will try toavoid downloading original chunks of content, because of the3G costs. To enforce cooperation and fairness among selfishpeers, incentive and scheduling are needed. In our system,every vehicle is connected to LTE network, so it behooves us toutilize a server assisted incentive scheme, which is much easierto implement and maintain than a fully distributed scheme.Hence we propose a server assisted key management scheme,where the server maintains the peer organization centrally,but the data content is distributed through low-cost V2V978-1-4673-7306-7/15/$31.00 c©2015 IEEE

transmissions.

Most existing approaches that improve the efficiency ofcontent downloading in vehicular networks, such as [2], [3],[4], [5], [6], [7], [1], [8], [9], rely on the cooperation amongvehicles. However, misbehavior among the vehicle communitymotivated by selfish intentions of saving 3G/LTE bandwidthor energy can severely degrade network performance. Hencemechanisms are needed to enforce the cooperation amongnodes in vehicular networks. Incentive schemes, such as [10],[11], [12], [13], [14], [15], [16], etc., are proposed to enforcethe cooperation of intermediate nodes to forward packetsin VANETs or other wireless networks. Generally, incentiveschemes can be categorized into two types, i.e., credit basedschemes and reputation based schemes. Credit based schemesare adopted by [10], [12], [13], [14], [16] where monetaryrewards/credits/tokens are allocated/paid to the forwarders. Areputation based scheme is adopted in [15], where users’behavior is recorded by reputation, and reward/punishmentis given according to the reputation score. A three-counterscheme was proposed in [11] where nodes count the number ofpackets requested to be forwarded and the number of packetsactually requested as a way of observing and recording thebehavior of neighbor nodes. This scheme is fairly similar tothe way how reputation schemes work. Table I summarizes therelated work according to the application scenarios, the use ofincentives and the type of incentive schemes used.

Power supply in vehicles is not a critical issue, hencenodes in VANETs usually do not have strong motivations todrop (instead of forwarding) packets to save energy. However,3G/LTE bandwidth is scarce for smart phones as well as smartvehicles. While some smart vehicles download interestingcontents from the Internet via 3G/LTE and share with theneighborhood in V2V manner, selfish vehicles try to be free-riders to benefit from contributing nodes without contributingthemselves in order to save 3G/LTE cost.

So far, no incentive schemes have been proposed to enforcecooperative, fair LTE downloading. In this paper, we proposean incentive compatible scheme for cooperative downloadingin vehicular networks. We first build a game theoretic model ofthe cluster downloading and sharing. Then we design a server-assisted round robin incentive and scheduling system usingkey management. Basically, the server encodes the sharedcontents with a temporary key and periodically updates it. Thescheme rewards contributive vehicles by providing them withthe keys to decode the content for the next few periods. Non-contributive nodes cannot decode until serving as cluster headsfor a certain amount of time. Since the wireless environmentof a vehicular network is highly lossy, and network codingcan significantly improve the reliability of packet deliveryaccording to existing research [17], we consider that networkcoding is applied for V2V packet sharing in our scheme.

In summary, the main contributions of our paper are:

• cluster-based content downloading and distribution forVANETs, leveraging both vechicle-to-infrastructure(V2I) and V2V communication in vehicular networksto save LTE bandwidth and cost.

• The cluster-based scheme is modeled as a gameand analyzed from a game theoretic perspective with

TABLE I. RELATED WORKS ON COOPERATIVE DOWNLOADING ANDINCENTIVE DESIGN IN VEHICULAR NETWORKS

Literatures Scenario Incentive type[2], [4], [5], [6], [1], [8], [9] Download None

[10], [12], [13], [14], [16] Forward Token/Credit[11], [15] Forward Counters/Reputation

Our scheme Download Key management

numerical analysis, proving that the cluster mem-ber vehicles cannot cooperate with efficiency andfairness without an incentive-compatible manage-ment/scheduling scheme.

• novel server-assisted key management scheme that en-forces the cooperation, fairness and security of clusterbased content downloading. Game theoretic analysisis done to verify cooperation enforcement and analyzesystem utility.

• simulations of a highway scenario with the proposedcluster-based content distribution as well as the keymanagement scheme. Results on selfish behavior im-pacts, system convergence, number of cluster headsand packet loss rate are presented and analyzed.

II. SYSTEM MODEL

A. Network model

We consider a vehicular network where vehicles the sameinterest form clusters, as shown in Fig. 2. A few cluster headsare selected and download the original data contents fromthe Internet source via 3G/LTE network. The cluster headsshare their data packets with member vehicles in the cluster,and members may share with each other in a peer-to-peerfashion. Since the highly mobile vehicular network usuallyhas a significant packet loss rate, in order to cope with the linkloss, the cluster heads may find it beneficial to perform networkcoding (with or without redundancy) before broadcasting thedata. A vehicle then can receive network coded packets fromdifferent cluster heads engaged in the same file download, andcan still decode and reconstruct the original content.

B. Game setting

We model the cluster-based content downloading and peer-to-peer distribution of each network coding generation as aone-shot cluster game G = (V,A, u). V = {v1, ..., vM}

Fig. 2. Cluster-based content distribution in vehicular networks

is the set of players, i.e., the set of vehicles in a cluster,where M denotes the total number of vehicles in a cluster.A = {a1, a2, ..., aM} is the set of actions taken by each vehiclevi, and ai ∈ {Head,Edge}, i.e., each vehicle can choose to beeither a cluster head or an edge node for the current round. Letm denote the number of vehicles choosing to be a cluster head,then the value of m represents the outcome of the actions. Theutility of a player vi, denoted as ui, can be one of the twovalues, depending on the action set. The player receives anexpected utility of uH(m), which is a constant as shown in(1), if he chooses to be a cluster head; on the other hand, itreceives an expected utility of uE(m), as shown in (2), if hechooses to be an edge node. We assume each node receives aconstant benefit of b for receiving each packet. The cluster headnodes encounter a constant cost of c for downloading eachpacket using 3G/LTE connection, and a constant cost of c′ forperforming network coding and broadcasting to other clustermembers. Since energy is not a critical issue for vehicles, thecost c′ is insignificant and ignored in the subsequent content.Should the model and solution in this paper be extendedto MANETs where energy is an important issue, c′ shouldbe added to c to indicate the cost of network coding andforwarding.

uH(m) = Kb−Kc (1)uE(m) = Kb · P (r ≥ K|m) (2)

where

P (r ≥ K|m) =

{∑mnr=K

(mnr

)(1− p)rp(mn−r), m ≥ 1

0, m = 0(3)

Here p denotes the link loss rate; K denotes the networkcoding generation size; n/K denotes the NC redundancyrate; and r denotes the number of packets actually received.P (r ≥ K) is the probability that an edge node receivesenough encoded packets (at least K) to decode and reconstructthe original data. If a node receives less than K packets fora generation, it cannot decode and receives 0 utility. Herewe assume the loss of each packet is independent, so thenumber of packets received follows a binomial distributionB(mn, 1 − p). In future work, we will take burst loss andother loss models into consideration. For multimedia contents,we propose that packets belonging to each I-frame are encodedas one NC generation with redundancy at cluster heads, whichserves as a “multi-path compatible network layer FEC”. Thisproposition takes advantage of the following three facts aboutvideo coding: (1) I-frames are large, thus have to be cut intomultiple packets; (2) I-frames are specifically important tothe video quality and need special protection, because theremaining frames in the group of pictures depend on them; and(3) generally, a frame cannot be decoded if one or more packetsare lost. The third fact about video coding exactly matcheswith the characteristic of network coding, hence encoding theframe in the network layer at cluster heads will provide ahigher reliability and decrease the chance of losing the wholegeneration/frame when less than K packets are received.

Table II gives an example of a 2-player game, i.e., assumingthere are two vehicles in the cluster v1 and v2. uH(m) isalways a constant uH ; and uE(m) depends on the number ofplayers choosing to be a cluster head.

TABLE II. THE UTILITY MATRIX OF 2-PLAYER CLUSTER GAME

v2Head Edge

v1Head uH , uH uH , uE(m = 1)Edge uE(m = 1), uH 0,0

C. Analysis of the one-shot cluster game

In order to analyze the cluster game, we need to find theNash equilibrium of the m-player game. Below is a definitionof Nash equilibrium.

Definition 1: Nash equilibrium for a strategic game is aprofile of actions such that each action is a best response tothe other actions. ai ∈ Ai is a best response to a−i ∈ A−i if

ui(ai, a−i) ≥ ui(a′i, a−i),∀a′i ∈ Ai (4)

where ai ∈ Ai is the i-th user’s action and a−i ∈ A−i denotesthe action profile of all users except user i.

To compute Nash equilibrium, we need to know whichis larger between uH(m) and uE(m). Given a fixed set ofNC generation size K, redundancy rate n/K, benefit and costscalers b and c, according to (1), uH is a constant, i.e., a clusterhead always has a fixed utility for downloading packets fromLTE. According to (2), uE(m) is an increasing function overthe number of cluster heads m, and a decreasing function overthe V2V link loss rate p. Fig. 3 shows how uE(m) variesover different link loss rates p comparing with uH ; Fig. 4shows how uE(m) varies with the number of cluster headsm comparing with uH . From the figures we can see that formost cases where the link loss rate is acceptable and there isat least one cluster head, the utility of being an edge node isalmost twice as much as being a cluster head, i.e., uE ' 2uH ,assuming the cost benefit ratio c/b is 0.5. Generally speaking,when the link loss rate is low, uE approaches b

b−cuH . Onlywhen the V2V link loss rate is so high that almost no packetscan get through, the utility of being a cluster head could behigher than being an edge node, i.e., uH > uE , because edgenodes receive almost nothing. One special case is that nobodyis willing to be cluster head, and everyone receives 0 utility,shown as the origin point in Fig. 4.

For this game, when uH ≥ uE(M − 1), e.g., whenpacket loss rate p = 0.95 in Fig. 4, there is one Nashequilibrium: ai = Head,∀vi ∈ V . Practically it refers to thesituation where the utility of being a cluster head is larger thanbeing an edge node, even if everyone else chooses to be a

Fig. 3. How the cluster head and edge node utilities vary with respect toV2V link loss rate. Settings: K = 8; n = 10; M = 100; b = 80; c = 40

Fig. 4. How the cluster head and edge node utilities vary with respect tothe number of cluster heads. Settings: K = 8; n = 10; M = 15; b = 80;c = 40

cluster head. This situation occurs when V2V communicationis too lossy to transmit data effectively. In this case, allnodes prefer to directly connect to the 3G/LTE and downloaddata by themselves. As system designers we do not need toworry about the cooperation enforcement in this extreme casebecause cooperation is not beneficial.

When uH < uE(M−1), there is always a m∗ ∈ (0,M−1]such that uE(m∗ − 1) < uH ≤ uE(m

∗), because uE(m) isan increasing function over m, and uH > uE(0) = 0. In thiscase, it forms a Nash equilibrium when exactly m∗ playerschoose to be a cluster head, and M −m∗ players choose tobe an edge node. m = m∗ is a Nash equilibrium because itwould result in a decrease of utility if any of the M − m∗

edge nodes deviates and changes its action to being a clusterhead since uE(m∗) ≥ uH ; it would also result in a decreaseof utility if any of the m∗ cluster heads deviates and changesits action to being an edge node, because uE(m∗ − 1) < uH .

The game has a Nash equilibrium that already maximizesthe overall utility, however, if we think practically, the Nashequilibrium cannot be efficiently achieved without the help ofan incentive compatible scheduling mechanism. If we assumenodes take actions all at the same time without negotiation,they would select the actions that maximize their expectedutilities according to their prior believes about m. If the nodesbelieve m ≥ m∗, all of them would choose to be edge nodes,leading to 0 utility for all nodes; if they believe m < m∗,all of them would choose to be cluster heads, resulting in nocooperation and a waste of resource. On the other hand, ifwe assume nodes choose their actions one by one, and thelater ones can observe what the others have chosen, then theNash equilibrium will be achieved when the first M − m∗

nodes choose to be edge nodes and the remaining peers have tobecome cluster heads, which is unfair and inefficient. Thereforeunder either assumption, the cluster-based scheme cannot workfairly and efficiently without a scheduling system, and itmust be incentive compatible such that selfish nodes wouldbe willing to follow the rule. Hence we propose a server-assisted management system to enforce efficient cooperationand fairness.

III. SERVER-ASSISTED KEY MANAGEMENT

Different from pure mobile ad hoc networks where centralauthority is not available, vehicular network is a mixed network

Fig. 5. Server-assisted key management system architecture

that contains both V2I and V2V communications. Moreover,given the security and privacy concerns, maintaining the vehi-cle cooperation in a pure distribution way (i.e., by broadcastingthe historical behavioral records among peers) is not feasibleor efficient. Hence, we choose to let V2I communications toconduct the task of management.

Given the broadcast nature of wireless networks, especiallyfor content distribution applications, nodes can always over-hear the data content, even without making any contributionover time. Also, security and privacy has always been a veryimportant concern for vehicular networks. Users may not wantto be overheard by unknown vehicles unless those vehicles areinterested in the same content and could be potential peers.Hence, to restrict the content distribution and keep it onlyamong verified contributive users, we propose a server or proxyassisted key management scheme.

A. How the key management scheme works

The server-assisted key management scheme, shown asFig. 5, basically works as follows. The server receives thecontent from the content source and encrypts the raw content,such as a video stream, with a temporary key. The key isupdated periodically. The server rewards the cluster heads bygiving them the keys for the next several periods, say P , afterthey have been serving as cluster head for a period. So forP periods, the previous cluster heads will be able to decodethe content. Those who have not recently been a cluster head(and therefore do not have the temporary key) are not able todecode, even if they overhear the V2V broadcast. They mustvolunteer to serve as a cluster head in order to receive the keyfrom the server.

The functions and operations includes the following:

1) Vehicles register with the server and create an accountthrough an application in order to participate, or login if they have one.

2) Vehicles submit their interest (say soccer gamestream) and location to the server. The server checksif there’s an existing download group (or cluster) inthe nearby area for the requested content. If yes, theserver adds the user to the group; otherwise the servercreates a new group for the user, initiates the contentdownloading from the original Internet source andencrypts with the key for the current time periodK(t).

Note that the range of the area does not have to bespecifically small or accurate, because as the vehiclestravel, they may meet other vehicles interested inthe same content that used to be quite far away10 minutes ago. For such cases, they don’t haveto leave the old group and join a different one,instead, they can seamlessly continue the cooperativedownload. In future work, we will investigate howlocation information can be used to improve the keymanagement scheme and the cluster head selectionalgorithm described in Section III.B. Note that allkeys are identical across all servers and clusters forthe same download group.

3) The server picks m (m ≥ m∗ according to SectionIV.A) vehicles to be cluster heads based on the clusterhead selection algorithm described in Section III.B,and starts sending the encrypted content. At the endof the period, it distributes the keys for the current andnext M/m periods to the cluster heads as rewards.

4) The server changes the keys and selects the newcluster heads for the next period. Step 3) and 4) arerepeated until the content download is completed.Upon completion, the group is dismissed and therecords are deleted.

B. Cluster head selection algorithm

We propose the following cluster head selection algorithm,where the server (or cluster manager) selects nodes that do nothave keys to decode with priority in order to promote fairness.

• A list is maintained at the server recoding which nodeshave been cluster heads in previous rounds, noted asH, which is a subset of all vehicles in the cluster V .

• At the beginning of each round, m vehicles areselected to be the cluster heads of the current roundHc. These m vehicles are randomly selected amongthe vehicles that are not in the list i.e., v ∈ V−H. Thenthe m nodes are added to the list, i.e., H ← H+Hc

• If the number of candidates is less than m, i.e., |V −H| < m, all nodes in |V − H| are selected as clusterheads for this round, and the rest m−|V −H| clusterheads are chosen from H. Then the list is cleared.

• Newly registered vehicles will be added to the memberset V , thus will have equal chance to be selected asthe others who have not been selected before. If somevehicles leave the cluster, the server will remove themfrom V .

It is efficient to prevent white wash (repeatedly leave andjoin the system to refresh records) by not granting the new-comer nodes the key to decode until they serve as cluster headfor the first time. Clearly, new users that just join the groupwill not be able to benefit until they contribute, therefore theyshould have strong intention to contribute instead of being free-riders. With this key management scheme, there is a decreasein throughput at the beginning since many nodes do not havekeys to decode the content they receive. Such decrease inthroughput serves as a tradeoff for the gain of cooperation, pri-vacy and security enforcement. Furthermore, the managementprocess incurs a cost of the server (thus the service provider),

however, the cooperative downloading scheme significantlysaves LTE bandwidth and reduces LTE network congestions,which is beneficial to the service provider.

IV. GAME THEORETIC ANALYSIS

A. Cooperation enforcement condition

In Section II.C, we showed that cluster-based content dis-tribution cannot be implemented fairly and efficiently withouta management/scheduling scheme. With the key managementscheme and corresponding cluster head selection algorithmpresented above, fairness and efficiency of cluster head se-lection is ensured. In this section, we use game theoretictools to analyze the cooperation enforcement when the keymanagement scheme is applied.

The game theoretic analysis is conducted based on the userinteraction per bMm c to dMm e periods, within which each clustermember serves as cluster head one time. The game is repeatedover time, however here we do not consider adding a discountfactor 1 to make it a repeated game, because the rewards (keys)are consumed within the M

m periods and do not affect futuredecisions. In other words, the user interactions of every M

mperiods are independent.

Let uC denote the expected utility of compliance, i.e.,serving as a cluster head for every bMm c to dMm e periods. LetuN denote the expected utility of non-compliance. A rationalnode will choose to comply if and only if uC ≥ uN . uC anduN can be expressed with the utility of being a cluster headuH(m), shown in (1) and the utility of being an edge nodeuE(m), shown in (2).

uC(m) = uH(m) + uE(m) · (Mm− 1) (5)

uN (m) = uH(m) · Mm

(6)

Since nodes that do not serve as cluster heads will not beable to receive the keys to decode, they have to download thewhole file on their own (if they are actually interested in thecontent), which results in a utility that is equivalent to beinga cluster head all the time. 2

The necessary and sufficient condition of compliance forany rational and selfish node is: uC(m) ≥ uN (m). Afterderivation, it is equivalent to: uE(m) ≥ uH(m).

According to Fig. 3 and the analysis in III.C, uE(m) ≥uH(m) is satisfied when the number of cluster head m is noless than a threshold m∗ = τ(p), which is an increasing func-tion of packet loss rate p, i.e., higher packet loss rate requiresmore cluster heads. When p approaches 1, m∗ approaches M ,which refers to the extreme case where everyone has to down-load by themselves (however the size of the cluster M does notaffect cooperation enforcement condition); when p approaches0, m∗ approaches 0 but m must be at least 1, which represents

1Discount factor in a repeated game represents the probability that the gamewill be repeated for an additional round.

2Non-compliant nodes download the content from a different source insteadof the cluster server, so they are not actually serving as cluster heads. Onlytheir payoffs/utilities are the same as cluster heads.

Fig. 6. How average utility varies with respect to the number of cluster head.Settings: K = 8; n = 10; M = 15; b = 80; c = 40

the best case where one cluster head is enough to serve a fleet.Hence, generally speaking, cooperation enforcement conditioncan be interpreted as m ≥ m∗ = τ(p). Evidently, the morecluster heads, the higher cooperation enforcement. With intra-flow network coding applied, the more cluster heads also endup with higher throughput and reliability.

B. Average utility

Let u denote the average utility of the cluster, defined tobe the average of all the nodes’ utilities in the cluster.

u(m) =

{mM · uH(m) + (1− m

M ) · uE(m), m ≥ 1

uH(m), m = 0(7)

If m < m∗, then uE(m) < uH(m) and u is increasing inm. This is because when there are too few cluster heads, edgenodes cannot receive sufficient packets. On the other hand,if m ≥ m∗, then uE(m) > uH(m) and u is decreasing inm. This is because when there are too many cluster heads,the benefit of cooperation decreases (to the non-cooperativecase for m = M ), which is very intuitive. Fig. 6 showshow average utility varies over number of cluster heads givendifferent packet loss rates. From the figure we can see thatthere is always a peak point, which is m∗. According to SectionIV.A, m∗ = τ(p). So the peak comes earlier when the packetloss rate p is small, i.e., when V2V communication is reliable,fewer cluster heads are needed; and the peak comes later whenp is large, i.e., when V2V communication is very lossy, morecluster heads are required.

C. Optimization & performance evaluation

Comparing the conclusions from Section IV.A and IV.B, wesee that there is an interesting relationship between the levelof cooperation and the average utility. In order to maximizethe average utility, yet enforce cooperation, we formalize theproblem as an optimization problem with constraint:

Maximize: u(m) =m

M· uH(m) + (1− m

M) · uE(m)

subject to: uE(m) ≥ uH(m)

The optimization problem is equivalent to finding theminimum number of cluster heads to enforce cooperation, andthe solution is clear, i.e., m∗. According to the model provided

in Section II, m∗ can be obtained by solving the equationuH(m) = uE(m) with a set of parameters (p, M , b, c) given,shown as the intersection points in Fig. 4 and the peak pointsin Fig. 6. Fig. 6 also shows that with a properly selectednumber of cluster heads, the performance of the cluster basedcontent downloading scheme, taking the average utility as theevaluation metric, is always better than the traditional situationwhere everybody downloads by themselves, represented by thehorizontal line in Fig. 6.

V. SIMULATION

As shown in the previous section, with a properly selectednumber of cluster heads, the performance of the cluster basedcontent downloading scheme is always better than the tradi-tional situation where everybody downloads by themselves.The proposed cluster-based downloading scheme with keymanagement is generally suitable for a stable group of usersin any type of wireless networks. As mentioned in the in-troduction section, one feasible application is video streamdownloading in vehicular ad hoc networks, especially highwayscenarios. To bring the analysis one step closer to reality,we performed a simulation on a particular VANET highwayscenario to

• study the impacts of selfish behaviors;

• verify the minimum number of cluster heads needed;

• study the impacts of the key management, i.e., the costof (or the performance degradation due to) incentives.

A. Simulation setup

We simulated the VANET highway scenario using SUMOand NS3. The simulation follows the system architectureshown in Fig. 5. The highway segment has a total lengthof 200 meters and contains 3 lanes and 4 exits. 50 vehicleswith the same interest move in the same direction at a startingvelocity of 60 mph and accelerate afterwards with randomness.On average 10% of the total traffic (i.e., 5 vehicles) will exitthe highway at each exit. The others remain on the highwayto the end. We did not set up highway entrances because thereare no ramp settings in SUMO that allow entering vehiclesto merge into the traffic; instead, both entering and oncomingvehicles must stop and wait as if it were an urban intersection,thus altering flow dynamics.

In the simulation, for simplicity we assume an ON/OFFpropagation loss model where receivers at or within themaximum propagation range receive the transmission at thetransmit power level, and receivers beyond the maximum rangereceive at power -1000 dBm (effectively zero). The valueof the maximum propagation range for V2V communicationimpacts the number of edge nodes within the range of eachcluster head. Equivalently, it impacts the number of clusterheads within the range of each edge node. When an edgenode is not within the range of a cluster head, the packetsbroadcasted by the cluster head won’t be received by the edgenode, equivalently lost. Hence, the propagation range impactspacket delivery. Fig. 7 shows how packet loss rate varies overpropagation range — larger propagation range leads to lowerpacket loss rate, which is intuitive. Of course, the lost packetsfrom one cluster head may be received from a different cluster

Fig. 7. Packet loss rate vs. maximum propagation range

head, so the packet loss rate here does not reflect final packetdelivery.

For the rest of the simulation, the propagation range is setto 30 meters, which results in an average packet loss rate ofp ' 22%. According to Fig. 6, when the loss rate is close to20%, the optimal number of cluster head is 2, which will beverified with the simulation result in part B.(1).

Each vehicle has a perfect LTE connection directly tothe cluster manager node (acting as the remote server). Thenode operations of packet downloading, distributions and keymanagement follows the descriptions in section III.

B. Simulation results

1) Non-compliant behavior impacts and the number ofcluster heads: We consider a probability of being non-compliant for each node in the cluster (equal to the fractionof non-compliant nodes in the system). When the serverdesignates a cluster member to be a cluster head, the nodebehaves as non-compliant, i.e., it rejects the task, with aprobability pN . Fig. 8 reports simulation results for packetreceiving rate and decoding rate by the edge nodes, i.e. the ratioof the number of packets received/decoded over the number ofpackets requested, as a function of non-compliance probabilitypN . The number of packets decoded is less than the numberof packets received by edge nodes because edge nodes whichhave not been a cluster head before do not have the keys todecode (as further discussed in part 2 of this section). Herewe set the number of cluster heads to 20, which means foreach requested packet, each edge node will ideally receive 20packets if it is within the range of all the cluster heads. Anincrease in non-compliance rate is equivalent to a decreaseof the number of effective cluster heads, i.e., a 95% non-compliance rate of 20 cluster heads is equivalent to havingonly 1 effective cluster head (the other 19 are non-compliant).

From the result we can see: (1) a higher compliance rate,i.e., a larger number of effective cluster heads, results in higherthroughput as well as higher packet loss tolerance; (2) for thescenario we simulated, two cluster heads, i.e. non-compliancerate = 0.9, can on average guarantee that the average deliverrate is above one. It verifies the result in Fig. 6 on the minimumnumber of cluster heads needed.

Without the stimulus of an incentive (and punishment),non-compliance behavior would be pervasive, considering thehuman nature of being a free-rider. The proportion of “al-truistic” users willing to pay for LTE service for the entiredownload is expected to be very small. For instance, with theoptimistic estimate of 5% altruistic users, when the system

Fig. 8. Packet receiving rate and decoding rate by edge nodes vs. non-compliance probability

Fig. 9. Decoding rate (cumulative average and per round) vs. number ofrounds

selects 20 cluster heads, there will be on average only onealtruistic cluster head responding, not sufficient (as per Fig. 8)to deliver the packets to all the member in the cluster for thehighway scenario we simulated. However with the proposedkey management scheme, we proved in Section IV.A that aslong as the number of selected cluster heads is no less than m∗(in the simulated scenario with p = 22%, m∗ = 2), the systemis incentive-compatible. This means that by selecting two ormore cluster heads, selfish users will have no motivation torefuse compliance because compliance yields a higher utilitythan non-compliance.

2) System convergence: We will use the following defini-tion as a metric for system convergence.

Definition 2: Let nd denote the number of packets receivedby edge nodes who have the key to decode, i.e., the decodablepackets. Let nt denote the total number of packets received byall edge nodes, with or without the key to decode. Decodingrate γ is defined to be the ratio of nd over nt.

γ =ndnt

(8)

As discussed at the end of Section III, since only a fewvehicles have the key to decode at the beginning phase,the throughput initially may be low, which is in fact thecost of cooperation enforcement (as well as security/privacyprotection). Fig. 9 shows the decoding rate (as defined above)of edge nodes over time, both cumulative and per round. Wecan see that the decoding rate increases very fast in the first afew rounds, and remains very close to 1 starting from round-7. The cumulative average reaches 1 only asymptotically as itmust aggregate all the previous rounds.

VI. CONCLUSION

In this paper, we propose a cluster-based scheme forcontent distribution in VANETs leveraging both V2I and V2Vcommunications. By forming a cluster, a few nodes actingas cluster heads download original content from the Internetvia 3G/LTE connections, and share with others in a peer-to-peer fashion via V2V communications. With game theoretictools and numerical analysis, we show that vehicle peerscannot spontaneously cooperate with efficiency and fairness.Hence we propose a server-assisted key management schemeto manage the cluster, enforcing cooperation and promotingsecurity and efficiency. We also propose a cluster head se-lection algorithm that ensures fairness across the peers. Bygame theoretic analysis and simulation of a highway scenario,we show that: (1) the system throughput increases with thenumber of effective cluster heads; (2) non-compliance behaviordecreases the number of effective cluster heads drastically, andthus reduces throughput; (3) our scheme converges very rapidlyafter a “slow-start” of decoding rate at initialization.

In future work, we will investigate how location infor-mation can be used to improve the key management schemeand the cluster head selection algorithm. For instance, packetloss rate could be reduced and key distribution could besped up by aggregating users into clusters based not only oninterests, but also on their geographic locations and relativeproximities. These clusters could be managed in the sameway as discussed in this paper, i.e., they could be split andcombined dynamically as proximity changes over time.

REFERENCES

[1] A. Nandan, S. Das, G. Pau, M. Gerla, and M. Sanadidi, “Co-operativedownloading in vehicular ad-hoc wireless networks,” in Wireless On-demand Network Systems and Services, 2005. WONS 2005. SecondAnnual Conference on, Jan 2005, pp. 32–41.

[2] M. Fiore and J. Barcelo-Ordinas, “Cooperative download in urbanvehicular networks,” in Mobile Adhoc and Sensor Systems, 2009. MASS’09. IEEE 6th International Conference on, Oct 2009, pp. 20–29.

[3] M. Gerla, C. Wu, G. Pau, and X. Zhu, “Content distribution in vanets,”Vehicular Communications, vol. 1, no. 1, pp. 3 – 12, 2014.

[4] H.-C. Jang and T.-Y. Hsu, “Urban multi-layered chord for p2p overvehicular network,” in Mobile and Wireless Networking (iCOST), 2012International Conference on Selected Topics in, 2012, pp. 54–59.

[5] M. Li, Z. Yang, and W. Lou, “Codeon: Cooperative popular contentdistribution for vehicular networks using symbol level network coding,”Selected Areas in Communications, IEEE Journal on, vol. 29, no. 1, pp.223–235, 2011.

[6] T. Luan, L. Cai, J. Chen, X. Shen, and F. Bai, “Engineering a distributedinfrastructure for large-scale cost-effective content dissemination overurban vehicular networks,” pp. 1–1, 2013.

[7] F. Malandrino, C. Casetti, C. F. Chiasserini, C. Sommer, and F. Dressler,“Content downloading in vehicular networks: Bringing parked cars intothe picture,” in Personal Indoor and Mobile Radio Communications(PIMRC), 2012 IEEE 23rd International Symposium on, 2012, pp.1534–1539.

[8] T. Wang, L. Song, and Z. Han, “Collaborative data dissemination incognitive vanets with sensing-throughput tradeoff,” in Communicationsin China (ICCC), 2012 1st IEEE International Conference on, 2012,pp. 41–45.

[9] D. Zhang and C.-K. Yeo, “A cooperative content distribution systemfor vehicles,” in Global Telecommunications Conference (GLOBECOM2011), 2011 IEEE, 2011, pp. 1–6.

[10] T. Chen, L. Zhu, F. Wu, and S. Zhong, “Stimulating cooperation invehicular ad hoc networks: A coalitional game theoretic approach,”Vehicular Technology, IEEE Transactions on, vol. 60, no. 2, pp. 566–579, 2011.

[11] Y. H. Ho, A. H. Ho, G. L. Hamza-Lup, and K. A. Hua, “Cooperationenforcement in vehicular networks,” in Proceedings of the 2008 Inter-national Conference on Communication Theory, Reliability, and Qualityof Service, ser. CTRQ ’08. Washington, DC, USA: IEEE ComputerSociety, 2008, pp. 7–12.

[12] F. Li and J. Wu, “Frame: An innovative incentive scheme in vehicularnetworks,” in Communications, 2009. ICC ’09. IEEE InternationalConference on, 2009, pp. 1–6.

[13] N. Mastronarde, V. Patel, J. Xu, and M. Van Der Schaar, “Learning re-laying strategies in cellular d2d networks with token-based incentives,”in Globecom Workshops (GC Wkshps), 2013 IEEE, Dec 2013, pp. 163–169.

[14] Z. Wang and C. Chigan, “Countermeasure uncooperative behaviors withdynamic trust-token in vanets,” in Communications, 2007. ICC ’07.IEEE International Conference on, 2007, pp. 3959–3964.

[15] C. Wu, M. Gerla, and M. van der Schaar, “Social norm incentives forsecure network coding in manets,” in Network Coding (NetCod), 2012International Symposium on, 2012, pp. 179–184.

[16] J. Xu and M. Van Der Schaar, “Token system design for autonomicwireless relay networks,” Communications, IEEE Transactions on,vol. 61, no. 7, pp. 2924–2935, July 2013.

[17] S.-H. Lee, U. Lee, K.-W. Lee, and M. Gerla, “Content distribution invanets using network coding: The effect of disk i/o and processing o/h,”in Sensor, Mesh and Ad Hoc Communications and Networks, 2008.SECON ’08. 5th Annual IEEE Communications Society Conference on,June 2008, pp. 117–125.