!" #$%&'%& ()*+,"-,%. ,% /,&0$11'%& 2+&1),"&,3 $1(1$!&4D1,6'%7

!"#$% &"$'()U%+'$,)+-(- &(,./) III -$

1(-,+-rcuevas@it.uc3m.es

1+23(. 4,5267(I%)8+8"8$ I1D:; <$8=/,7) (%-U%+'$,)+-(- &(,./) III -$

1(-,+-michal.kryczka@imdea.org

;%>$. &"$'()U%+'$,)+-(- &(,./) III -$

1(-,+-acrumin@it.uc3m.es

?$#()8+(% 4("%$4@1 A(#

BU D(,C)8(-8kaune@kom.tu-darmstadt.de

&(,C$% D"$,,$,/U%+'$,)+-(- &(,./) III -$

1(-,+-guerrero@it.uc3m.es

!$6( !$E(+$!

U%+'$,)+85 /F @,$>/%reza@cs.uoregon.edu

ABSTRACTBitTorrent is the most popular P2P content delivery appli-cation where individual users share various type of contentwith tens of thousands of other users. The growing popular-ity of BitTorrent is primarily due to the availability of valu-able content without any cost for the consumers. However,apart from required resources, publishing (sharing) valuable(and often copyrighted) content has serious legal implica-tions for users who publish the material (or publishers). Thisraises a question that whether (at least major) content pub-lishers behave in an altruistic fashion or have other incen-tives such as !nancial. In this study, we identify the contentpublishers of more than 55K torrents in twomajor BitTorrentportals and examine their behavior. We demonstrate that asmall fraction of publishers is responsible for 67% of thepublished content and 75% of the downloads. Our investi-gations reveal that these major publishers respond to two dif-ferent pro!les. On the one hand, antipiracy agencies and ma-licious publishers publish a large amount of fake !les to pro-tect copyrighted content and spread malware respectively.On the other hand, content publishing in BitTorrent is largelydriven by companies with !nancial incentives. Therefore, ifthese companies lose their interest or are unable to publishcontent, BitTorrent traf!c/portals may disappear or at leasttheir associated traf!c will be signi!cantly reduced.!Reza Rejaie was Visiting Researcher at Institute IMDEANetworks, September 2009 - August 2010, while this studywas performed.

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for pro!t or commercial advantage and that copiesbear this notice and the full citation on the !rst page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior speci!cpermission and/or a fee.ACM CoNEXT 2010, November 30 – December 3 2010, Philadelphia,USA.Copyright 2010 ACM 1-4503-0448-1/10/11 ...$10.00.

Categories and Subject DescriptorsC.2.4 [Computer Communication Networks]: Dis-tributed Systems – Distributed Applications

General TermsMeasurement, Economics

KeywordsBitTorrent, Content Publishing, Business Model

1. INTRODUCTIONPeer to Peer (P2P) file-sharing applications, and more

specifically BitTorrent, are a clear example of killer In-ternet applications in the last decade. BitTorrent iscurrently used by hundreds of millions of users and isresponsible for a large portion of the Internet tra!c[4]. This in turn has attracted the research commu-nity to examine various aspects of swarming mechanismin BitTorrent [10, 16, 13] and propose di"erent tech-niques to improve its performance [19, 15]. Further-more, other aspects of BitTorrent such as demographyof users [22, 12, 20] along with the security [21] and pri-vacy [6, 7] issues have also been studied. However, thesocio-economic aspects of BitTorrent in particular, andother P2P file sharing systems in general, have receivedlittle attention. In particular, the availability of popu-lar and often copyrighted content (e.g. recent TV showsand Hollywood movies) to millions of interested users atno cost is the key factor in the popularity of BitTorrent.This raises an important question about the incentiveof publishers who make these content available throughBitTorrent portals. Despite its importance to properlyunderstand the popularity of BitTorrent, to our knowl-edge prior studies on BitTorrent have not tackled thiscritical question.

In this paper, we study content publishing in Bit-

Torrent from a socio-economic point of view by unrav-elling who publishes content in BitTorrent, and why.Toward this end, we conduct a large scale measurementover two major BitTorrent portals, namely Mininovaand the Pirate Bay, to capture more than 55K publishedcontent that involve more than 35M IP addresses. Us-ing this dataset, we first examine the contribution ofthe individual content publishers and illustrate that asmall fraction of publishers (∼100) are responsible ofuploading 67% of the content that serve 75% of thedownloads in our major dataset. Furthermore, mostof these major publishers dedicate their resources forpublishing content and consume few or no publishedcontent by others, i.e. their level of content publica-tion and consumption is very imbalanced. In additionto allocating a significant amount of resources for pub-lishing content, these users often publish copyrightedmaterial, which has legal implications for them [1, 2].These observations raise the following question: whatare the main incentives of (major) content publishersin BitTorrent?.

To answer this important question, we conduct a sys-tematic study on the major publishers in BitTorrent.We show that these publishers can be broadly dividedinto two di"erent groups: fake publishers who publisha large number of fake content and top publishers whopublish a large number of often copyrighted content.We also identify the main characteristics (i.e. signature)of publishers in each group such as their seeding behav-ior and the popularity of their published content. Weinvestigate the main incentives of major (non-fake) pub-lishers and classify them into three categories (i) Pri-vate BitTorrent Portals that o"er certain services andreceive financial gain through ads, donations and fees,(ii) Promoting web sites that leverage published con-tent at BitTorrent portals to attract users to their ownweb site for financial gain, and (iii) Altruistic MajorPublishers. We characterize these three groups of pub-lishers and present the estimated value (income) of theassociated web sites to support our claims about theirincentives.

The main contributions of this paper can be summa-rized as follows:

• We present a simple measurement methodology tomonitor the content publishing activity in majorBitTorrent portals. This methodology has beenused to implement a system that continuously mon-itors and reports the content publishing activity inthe Pirate Bay portal. The collected data is madepublicly available through a web site.

• The distribution of the number of published con-tent by each publisher is very skewed, i.e. a verysmall fraction of publishers (∼100) is responsiblefor a significant fraction of the published content

(67%) and even more significant fraction of thedownloads (75%). These major publishers can befurther divided into three groups based on their in-centives as follows: fake publishers, altruistic toppublishers and profit-driven publishers.

• Fake publishers are either antipiracy agencies ormalicious users who are responsible for 30% of thecontent and 25% of the downloads. These pub-lishers sustain a continuous poisoning-like indexattack [17] against BitTorrent portals that a"ectsmillions of downloaders.

• Profit-driven top publishers own fairly profitableweb sites. They use major BitTorrent portals suchas the Pirate Bay as a platform to advertise theirweb sites to millions of users. For this purposethey publish popular torrents where they attachthe URL of their web sites in various manners.The publishers that pursue this approach are re-sponsible for roughly 30% of the content and 40%of the downloads in BitTorrent.

The rest of the paper is organized as follows. Section2 describes our measurement methodology. Sections 3and 4 are dedicated to the identification of major pub-lishers and their main characteristics (i.e. signature) re-spectively. In Section 5, we study the incentives thatmajor publishers have to perform this activity. Section6 presents other players that also benefit from contentpublishing. In Section 7 we describe our publicly avail-able application to monitor content publishing activityin the Pirate Bay portal. Finally Section 8 discussesrelated work and Section 9 concludes the paper.

.2. MEASUREMENT METHODOLOGY

This section describes our methodology to identifythe initial publisher of a file that is distributed througha BitTorrent swarm. Towards this end, we first brieflydescribe the required background on how a user joins aBitTorrent swarm.Background: A BitTorrent client takes the followingsteps to join the swarm associated with file X . First, theclient obtains the .torrent file associated to the desiredswarm. The .torrent file contains contact informationfor the tracker that manages the swarm and the num-ber of pieces of file X . Second, the client connects tothe tracker and obtains the following information: (i)the number of seeders and leechers that are currentlyconnected to the swarm, and (ii) N (typically 50) ran-dom IP addresses of participating peers in the swarm.Furthermore, if the number of neighbors is eventuallylower than a given threshold (typically 20), the clientcontacts the tracker again to learn about other peers inthe swarm.

To facilitate the bootstrapping process, the .torrentfiles are typically indexed at BitTorrent portals. Some

Portal Start End #Torrents #IPaddresses

mn08 Mininova 09-Dec-08 16-Jan-09 - /20.8K 8.2Mpb09 Pirate Bay 28-Nov-09 18-Dec-09 23.2K/10.4K 52.9Kpb10 Pirate Bay 06-Apr-10 05-May-10 38.4K/14.6K 27.3M

Table 1: Datasets Description.

of the major portals (e.g. the Pirate Bay or Mininova1)index millions of .torrent files [22], classify them intodi"erent categories and provide a web page with de-tailed information (content category, publisher’s user-name, file size, and file description) for each file. Theseportals also o"er an RSS feed to announce a newly pub-lished file. The RSS feed provides some informationsuch as content category, content size and publisher’susername for a new file.Identifying Initial Publisher: The objective of ourmeasurement study is to determine the identity of theinitial publishers of a large number of torrents and to as-sess the popularity of each published file (i.e. the num-ber and identity of peers who download the file).

Toward this end, we leverage the RSS feed to de-tect the availability of a new file on major BitTorrentportals and retrieve the publisher’s username. In orderto obtain the publisher’s IP address, we immediatelydownload the .torrent file and connect to the associatedtracker. This implies that we often contact the trackershortly after the birth of the associated swarm whenthe number of participating peers is likely to be smalland the initial publisher (i.e. seeder) is one of them.We retrieve the IP address of all participating peers aswell as the current number of seeders in the swarm. Ifthere is only one seeder in the swarm and the num-ber of participating peers is not too large (i.e. < 20),we obtain the bitfield of available pieces at individualpeers to identify the seeder. Otherwise, reliably iden-tifying the initial seeder is di!cult because there aremore than one seeder or the number of participatingpeers is large2. Furthermore, we cannot directly con-tact the initial seeder that is behind a NAT box andthus we are unable to identify the initial publisher’s IPaddress in such cases. Using this technique we were ableto reliably identify the publisher’s username for all thetorrents and the publisher’s IP address in at least 40%of the torrents.

Once we identify a publisher, we periodically querythe tracker in order to obtain the IP addresses of theparticipants in the associated swarm and always solicitthe maximum number of IP addresses (i.e. 200) from

1http://thepiratebay.org/, http://www.mininova.org/.2Our investigations revealed two interesting scenarios forwhich we could not identify the initial publisher’s IP ad-dress: (i) swarms that have a large number of peers shortlyafter they are added to the portal. We discovered that theseswarms have already been published in other portals. (ii)swarms for which the tracker did not report any seeder fora while or did not report a seeder at all.

the tracker. To avoid being blacklisted by the tracker,we issue our queries at the maximum rate that is allowedby the tracker (i.e. 1 query every 10 to 15 minutes de-pending on the tracker load). Given this constraint, wequery the tracker from several geographically-distributedmachines so that the aggregated information by all thesemachines provides an adequately high resolution viewof participating peers and their evolution over time. Wecontinue to monitor a target swarm until we receive 10consecutive empty replies from the tracker. We use theMaxMind Database [3] to map all the IP addresses (forboth publishers and downloaders) to their correspond-ing ISPs and geographical locations.

2.1 DatasetUsing the described methodology, we identify a large

number of BitTorrent swarms at two major BitTorrentportals, namely Mininova and the Pirate Bay. Each oneof these portals was the most popular BitTorrent portalat the time of the corresponding measurement accord-ing to Alexa ranking3. It is worth noting that the PirateBay is in particular interesting for our study since it isthe only main BitTorrent portal where all the publishedcontent is contributed by users [22] (as opposed to be-ing retrieved from other portals). Table 1 shows themain features of our three datasets (1 from Mininovaand 2 from the Pirate Bay) including the start and enddates of our measurement, the number of torrents forwhich we identified the initial publisher (username/IPaddress), and the total number of discovered IP ad-dresses associated for all the monitored swarms. Werefer to these datasets as mn08, pb09 and pb10 through-out this paper. We note that dataset mn08 does notcontain the username of initial publishers, and we use asingle query to identify initial publishers in dataset pb09after detecting a new swarm through the RSS feed. Weuse all three datasets for our general analysis but fo-cused on pb10 for our detailed analysis.

3. IDENTIFYING MAJOR PUBLISHERSA publisher can be identified by its username and/or

IP address. In our analysis, we identify individual pub-lishers primarily by their username since the usernameis expected to remain consistent across di"erent tor-rents. The only exceptions are publishers in mn08 sincewe do not have their usernames and fake publishers aswe describe in the next subsections.

3.1 Skewness of ContributionFirst, we examine the level of contribution (i.e. the

number of published files) by the identified content pub-lishers in each dataset. Figure 1 depicts the percentageof files that are published by the top x% of the publish-ers in our three datasets. We observe that the top 3% of3http://www.alexa.com/topsites.

mn08 pb09 pb10

ISP Type % ISP Type % ISP Type %OVH Hosting Provider 13.31 OVH Hosting Provider 24.76 OVH Hosting Provider 15.16Comcast Commercial ISP 4.69 Comcast Commercial ISP 3.67 SoftLayer Tech. Hosting Provider 4.52Keyweb Hosting Provider 3.18 Road Runner Commercial ISP 2.3 FDCservers Hosting Provider 3.64Road Runner Commercial ISP 3.03 Romania DS Commercial ISP 2.27 Open Computer Network Commercial ISP 3.59NetDirect Hosting Provider 2.44 MTT Network Commercial ISP 1.95 tzulo Hosting Provider 3.36Virgin Media Commercial ISP 2.42 Verizon Commercial ISP 1.64 Comcast Commercial ISP 2.86NetWork Operations Center Hosting Provider 2.39 Virgin Media Commercial ISP 1.49 Cosema Commercial ISP 2.25SBC Commercial ISP 2.38 SBC Commercial ISP 1.41 Telefonica Commercial ISP 2.22Comcor-TV Commercial ISP 2.33 NIB Commercial ISP 1.26 Jazz Telecom. Commercial ISP 2.07Telecom Italia Commercial ISP 2.02 tzulo Hosting Provider 1.14 4RWEB Hosting Provider 2.06

Table 2: Content Publishers Distribution per ISP.

G 2G IG JG KG 1GGG

2G

IG

JG

KG

1GG

M$,2$%8(>$N/FNNO"#.+)3$,)

M$,2$%8(>$N/FNO"#.+)3$-N2/%8$%8

N

N

C%GKO#GPO#1G

Figure 1: Percentage of content published by thetop x% publishers.

the BitTorrent publishers contribute roughly 40% of thepublished content. Moreover, a careful examination ofIP addresses for the top-100 (i.e. 3%) publishers in ourpb10 dataset reveals that a significant fraction of themeither do not download any content (40%) or downloadless than 5 files (80%). This large contribution of re-sources (bandwidth and content) by major publisherscoupled with the significant imbalance between theirpublishing and consumption rates seems non-altruisticand rather di!cult to justify for two simple reasons:

- Required Resources/Cost : publishing a large num-ber of content requires a significant amount of pro-cessing and bandwidth. For example, a major con-tent publisher named eztv recommends in its privateBitTorrent portal web page (www.eztv.it) to allocateat least 10Mbps in order to sustain the seeding of few(around 5) files in parallel.- Legal Implications: As other studies have reported[6] and we confirm in our datasets, a large fraction ofcontent published by major publishers is copyrightedmaterial (recent movies or TV series). Thus, publishingthese files is likely to have serious legal consequences forthese publishers [1, 2].

This raises the question that why this small fractionof publishers allocate a great deal of (costly) resourcesto contribute many files into BitTorrent swarms despitepotential legal implications?. We answer this centralquestion in Section 5.

3.2 Publishers’ ISPsTo help identify content publishers in our dataset, we

determine the ISP that hosts each major publisher anduse that information to assess the type of service (andavailable resources) that a publisher is likely to have.Toward this end, we map the IP address for a publisherin each dataset to its corresponding ISP using the Max-Mind database [3]. We then examine publicly availableinformation about each ISP (e.g. its web page) to de-termine whether it is a commercial ISP or a hostingprovider. We perform this analysis only for the top-100 (roughly 3%) publishers since these publishers aremostly of interest and collecting the required informa-tion for all publishers is a tedious task. Since we donot have publishers’ username for mn08, we examinethe top-100 publishers based on their IP addresses inthis dataset. For these publishers, we cannot assess theaggregated contribution of a publisher through di"erentIP addresses (i.e. under-estimating the contribution ofeach publisher).

We observe that 42% of the top-100 publishers inpb10, 35% of the top-100 in pb09 and 77% of the top-100 publishers in mn08 are located at hosting services.Moreover 22%, 20% and 45% of these top-100 publish-ers are located at a particular hosting services, namelyOVH, in pb10, pb09 and mn08 respectively.

In short, our analysis reveals that a significant frac-tion of major publishers are located at a few hostingservices and a large percentage of them at OVH.

We also examine the contribution of BitTorrent pub-lishers at the ISP-level by mapping all the publishers totheir ISPs and identify the top-10 ISPs based on theiraggregate published content for each dataset as shownin Table 2. This table confirms that content publisherswho are located at a particular hosting provider, namelyOVH, have consistently contributed a significant frac-tion of published content at major BitTorrent portals.There are also several commercial ISPs (e.g. Comcast)in Table 2 with a much smaller contribution.

To assess the di"erence between users from hostingproviders and commercial ISPs, we compare and con-trast the characteristics of all publishers that are lo-cated at OVH and Comcast as representative ISPs foreach class of publishers in Table 3. This table demon-strates the following two important di"erences: first,

Published # IP addr # /16 IP # Geotorrents Pref. Loc.

OVH (mn08) 2766 164 5 2Comcast (mn08) 976 675 269 400OVH (pb09) 2577 78 5 2Comcast (pb09) 382 198 143 129OVH (pb10) 2213 92 7 4Comcast (pb10) 408 185 139 147

Table 3: Characteristics of all OVH and Com-cast publishers in mn08, pb09 and pb10.

the aggregate contribution of each publisher at OVH ison average a few times larger than Comcast publish-ers. Second, Comcast publishers are sparsely scatteredacross many /16 IP prefixes and many geographical lo-cations in the US whereas OVH publishers are concen-trated in a few /16 IP prefixes and a handful of di"er-ent locations in Europe (i.e. the location of OVH’s datacenters). In essence, the published content by Comcastpublishers comes from a large number of typical altru-istic users where each one publishes a small number offiles from their home or work. In contrast, OVH pub-lishers appear to be paying for a well provisioned serviceto be able to publish a much larger number of files. Wehave also examined consumer peers in captured torrentsand did not observe the presence of OVH users amongthe consuming peers.

In summary, the examination of ISPs that host majorBitTorrent publishers suggests that these publishers arelocated either at a few hosting providers (with a largeconcentration at OVH) or at commercial ISPs. Thesepublishers contribute a significantly larger number offiles than average publishers. Furthermore, publisherswho are located at hosting providers do not consumepublished content by other publishers.

3.3 A Closer Look at Major PublishersWe now examine the mapping between username and

IP address of the top-100 content publishers in the pb10dataset to gain some insight about major publishers be-havior. Our examination reveals the following interest-ing points:

First, if we focus on the top-100 IP addresses thathave published the largest number of files, only 55%of them are used by a unique username. The remain-ing 45% of the IP addresses of major publishers aremapped to a large number of usernames. We have care-fully investigated this set of IP addresses and discoveredthat they use either hacked or manually created ac-counts (with a random username) to inject ”fake” con-tent. These publishers appear to be associated with an-tipiracy agencies or malicious users. The former grouptries to avoid the distribution of copyrighted contentwhereas the latter attempts to disseminate malware.We refer to these publishers as fake publishers. Surpris-ingly, fake publishers are responsible for around 25%of the usernames, 30% of the published content and

25% of the downloads in our pb10 dataset. This sug-gests that major BitTorrent portals are su"ering from asystematic poisoning index attack [17] that a"ects 30%of the published content. The portals fight this phe-nomenon by removing the fake content as well as theuser accounts used to publish them. However, contraryto what has been reported in previous studies [20], thistechnique does not seem to be su!ciently e"ective sincemillions of users initiate the download of fake content.Finally, it is worth noting that most of the fake pub-lishers perform their activity from three specific host-ing providers named tzulo, FDC Servers and 4RWEB.Due to the relevant activity of these fake publishers, westudy them as a separate group in the rest of the paper.

Second, the inspection of the top-100 usernames whopublish the largest number of files shows that only 25%of them operate from a single IP. The remaining 75% oftop usernames utilize multiple IPs and can be classifiedinto the following common cases: (i) 34% of the user-names with multiple IP addresses (5.7 IP addresses onaverage) at a hosting provider in order to obtain the re-quired resources for seeding a large number of files. (ii)24% of the usernames with multiple IP addresses (13.8IP addresses on average) located at a single commer-cial ISP. Their mapping to multiple IP addresses mustbe due to the periodical change of their assigned IPaddresses by their ISPs. (iii) The other 17% of theseusernames are mapped to multiple IP addresses (7.7IP addresses on average) at di"erent commercial ISPs.These are users who inject content from various loca-tions (e.g. home and work computer). To properly char-acterize di"erent types of publishers, we exclude the 16usernames who publish fake content from the top-100usernames. We refer to the remaining top-100 user-names (non-fake publishers) as Top publishers who areresponsible of 37% of the published content and 50% ofthe total downloads in our pb10 dataset.

In summary, the major portion of the content comesfrom two reduced group of publishers: Top publishersand Fake publishers that collectively are responsible of67% of the published content and 75% of the downloads.In the rest of the paper we devote our e"ort to charac-terize these two groups.

4. SIGNATURE OFMAJOR PUBLISHERSBefore we investigate the incentives of major Bit-

Torrent publishers, we examine whether they exhibitany other distinguishing features, i.e. whether majorpublishers have a distinguishing signature. Any suchdistinguishing features could shed some light on theunderlying incentives of these publishers. Toward thisend, in the next few subsections, we examine the follow-ing characteristics of major publishers in our datasets:(i) the type of published content, (ii) the popularity ofpublished content, and (iii) the availability and seeding

(a) mn08 (b) pb10

Figure 2: Type of published content distributionfor the di!erent classes of publishers: All, Fake,Top, Top-HP and Top-CI.

behavior of a publisher.To identify distinguishing features, we examine the

above characteristics across the following three targetgroups in each dataset: all publishers (labeled as “All”),all fake publishers (labeled as “Fake”) and all top-100(non-fake) publishers regardless of their ISPs (labeledas “Top”). We also examine the break down of Toppublishers based on their ISPs into hosting providersand commercial ISPs, labeled as “Top-HP” and “Top-CI”, respectively.

4.1 Content TypeWe leverage the reported content type by each pub-

lisher to classify the published content across di"erenttarget groups. Figure 2 depicts the break down of pub-lished content across di"erent content type for all pub-lishers in each target group for our Mininova and ourmajor Pirate Bay datasets. We recall that without user-name information for each publisher in mn08 dataset,we cannot identify fake publishers. Figure 2 reveals afew interesting trends as follows:

First, Video files (which mainly include movies, TV-shows and porn content) constitute a significant frac-tion of published files across most groups with some im-portant di"erences. The percentage of published videoacross all publishers is around 37%-51% but it is slightlylarger among top publishers. However, video is clearlya larger fraction of published content by the top pub-lishers located at hosting providers in our pb10 dataset.Fake publishers primarily focus on Videos (recent moviesand TV shows) and Software content. This supports ourearlier observation that these publishers consist of an-tipiracy agencies and malicious users because the formergroup publishes a fake version of recent movies while thelatter provides software that contains malware.

4.2 Content PopularityThe number of published files by a publisher shows

only one dimension of its contribution to BitTorrent.The other equally important issue is the popularity ofeach published content (i.e. the number of download-

;.. Q(7$ B/O B/O!RM B/O!&I1G1

1G2

1G3

1GI

;'>N<"

CND/=

%./(-$,)NO$,NB/,,$%8NO$,NM"#.+)3$,

Figure 3: Avg Num of Downloaders per torrentper publisher for the di!erent classes of publish-ers: All, Fake, Top, Top-HP, Top-CI.

ers regardless of their download progress) by individualpublishers. Figure 3 shows the box plot of the distribu-tion of average downloaders per torrent per publisheracross all publishers in each target group where eachbox presents the 25th, 50th and 75th percentiles.

On the one hand, the median popularity of top pub-lishers’ torrents is 7 times higher than a typical user(represented by All). A closer examination of the Toppublishers shows that the content published by users athosting providers is on average 1.5 times more popularthan those published by users at commercial ISPs. Onthe other hand, fake publishers’ content is the most un-popular among the target groups. This is because theportals actively monitor the torrents and immediatelyremove the content identified as fake to avoid users fromdownloading it. Furthermore, users quickly realize thefake nature of these content and report this info on fo-rums that inform others and limit their popularity.

In summary, top publishers are responsible for a largerfraction of popular torrents. This in turn magnifies thecontribution of the 37% of the injected files by the toppublishers to be responsible for 50% of all the down-loads. The low popularity of fake publishers’ contenthas the opposite e"ect and limits their contribution tothe number of downloads to 25%.

4.3 Seeding BehaviorWe characterize the seeding behavior of individual

publishers in our target groups using the following met-rics: (i) average seeding time of a publisher for its pub-lished content, (ii) average number of parallel seededtorrents, and (iii) aggregated session time of a publisheracross all its torrents. Since calculating these propertiesrequires detailed analysis of our dataset that are com-putationally expensive, we are unable to derive thesevalues for all publishers. We use 400 randomly selectedpublishers to represent the normal behaviour of all pub-lishers and refer to this group as “All” in our analysis.

In order to compute these metrics we need to estimate

;.. Q(7$ B/O B/O!RM B/O!&IG

2G

IG

JG

KG

1GG;'>TN?$$-+%>NB+C$NM$,NB/,,$%8NM$,NM"#.+)3$,NU3/",)V

(a) Average Seeding Time per Contentper Publisher

;.. Q(7$ B/O B/O!RM B/O!&IG

W

1G

1W

2G

2W

3G

3W

IG

;'>N<"

CN/FNB/,,$%8)N?$$-N+%NM(,(..$.NO$,NM"#.+)3$,

(b) Average Number of Parallel SeedTorrents Per Publisher

;.. Q(7$ B/O B/O!RM B/O!&IG

1GG

2GG

3GG

IGG

WGG

JGG

;>>,$>(8$-N?$))+/

%NB+C$NO$,NM

"#.+)3$,NUR/

",)V

(c) Average Aggregated Session TimePer Publisher

Figure 4: Seeding Behaviour for the di!erent classes of publishers: All, Fake, Top, Top-HP andTop-CI.

the time that a specific publisher has been connected toa torrent (in one or multiple sessions). Since each queryto the tracker just reports (at most) a random subset of200 IPs, in big torrents (>200 peers), we need to per-form multiple queries in order to assess the presence ofthe publisher in the torrent. In our associated Techni-cal Report [8], we detail the technique used to estimatethe session time of a specific user in each torrent.Average Seeding Time: We measure the duration oftime that a publisher stays in a torrent since its birth toseed the content. In general, a publisher can leave thetorrent once there is an adequate fraction of other seeds.Figure 4(a) depicts the summary distribution of averageseeding time across all publishers in each target group.This figure demonstrates the following points: first, theseeding time for fake publishers is significantly longerthan publishers in other groups. Since these publishersdo not provide the actual content, the initial fake pub-lisher remains the only seed in the session (i.e. otherusers do not help in seeding fake content) to keep thetorrent alive. Second, Figure 4(a) shows that top pub-lishers typically seed a content for a few hours. How-ever, the seeding time for top publishers from hostingproviders is clearly longer than top publishers from com-mercial ISPs. This suggests that publishers at hostingproviders are more concerned about the availability oftheir published content.Average number of Parallel Torrents: Figure 4(b)depicts the summary distribution of the average num-ber of torrents that a publisher seeds in parallel acrosspublishers in each target group. This figure indicatesthat fake publishers seed many torrents in parallel. Wehave seen that fake publishers typically publish a largenumber of torrents and other users do not help them forseeding. Therefore, fake publishers need to seed all oftheir seeded torrents in parallel in order to keep themalive. The results for top publishers show that theirtypical number of seeded torrents in parallel is the same

(around 3 torrents) regardless of their location. How-ever, we expect that a regular publisher seed only 1 fileat a time.Aggregated Session Time: We have also quanti-fied the availability of individual publishers by estimat-ing the aggregated session time that each publisher ispresent in the system across all published torrents. Fig-ure 4(c) shows the distribution of this availability mea-sure across publishers in each target group. As expectedfake publishers present the longest aggregated sessiontime due to their obligation to continuously seed theircontent to keep them alive. If we focus on top publish-ers, they exhibit a typical aggregated session 10 timeslonger than standard users. Furthermore, top publish-ers at hosting services are clearly more available thanthose from commercial ISPs.

4.4 SummaryBitTorrent content publishers can be broadly divided

into three groups as follows: (i) Altruistic users whopublish content while consuming content that is pub-lished by other users. (ii) Fake publishers publish asignificant number of files that are often Video andSoftware content from a few hosting providers. Dueto the fake nature of their content, their torrents areunpopular and they need to seed all torrents to keepthem alive. These publishers appear to be associatedwith antipiracy agencies or malicious users. We vali-date this hypothesis in the next section. (iii) Top pub-lishers publish a large number of popular files (oftencopyrighted video) and remain for a long time in the as-sociated torrents to ensure proper seeding of their pub-lished content. These publishers are located at hostingfacilities or commercial ISPs. Their behavior suggeststhat these publishers are interested in the visibility ofthe published content possibly to attract a large num-ber of users. The (cost of) allocated resources by thesepublishers along with legal implications of publishingcopyrighted material cannot be considered as altruistic.

Therefore, the most conceivable incentive for these pub-lishers appears to be financial profit. We examine thishypothesis in the rest of the paper.

5. INCENTIVES OFMAJOR PUBLISHERSIn this section, we examine the incentive of di"er-

ent groups of publishers in more detail. First we fo-cus on fake publishers by examining the name and con-tent of the files they published. We noticed that thesepublishers often publish files with catchy titles (e.g. re-cently released Hollywood movies) and in most casesthe actual content has already been removed at themoment of downloading it4. The few files we wereable to download were indeed fake content. Some ofthem included an antipiracy message whereas some oth-ers led to malware software. In the latter case, thecontent was a video that had a pointer to a specificsoftware (e.g. http://flvdirect.com/) to be down-loaded in order to reproduce the video. This softwarewas indeed a malware5. These observations validateour hypothesis that fake publishers are either antipiracyagents who publish fake versions of copyrighted contentor malicious users who led users to download a malware.Therefore, we have clearly identified the incentives offake publishers.

Second, another group of major publishers allocatesignificant amount of resources to publish non-fake anoften copyrighted content. We believe that the behav-ior of these users is not altruistic. More specifically, ourhypothesis is that these publishers leverage major Bit-Torrent portals as a venue to freely attract downloadingusers to their web sites. To verify this hypothesis, weconduct an investigation to gather the following infor-mation about each one of the top (i.e. top-100 non-fake)publishers:

• Promoting URL: the URL that downloaders of apublished content may encounter,

• Publisher’s Username: any publicly available in-formation about the username that a major pub-lisher uses in the Pirate Bay portal, and

• Business Profile: o"ered services (and choices) atthe promoting URL.

Next, we describe our approach for collecting this infor-mation.Promoting URL: We emulate the experience of a userby downloading a few randomly-selected files publishedby each top publisher to determine whether and wherethey may encounter a promoting URL. We identified4We tried to download these files a few weeks after the cor-respondent measurement study was performed.5http://www.prevx.com/filenames/X2669713580830956212-X1/

FLVDIRECT.EXE.html

three places where publishers may embed a promotingURL: (i) name of the downloaded file (e.g. user mois20names his files as filename-divxatope.com, thus adver-tising the URL www.divxatope.com), (ii) the textboxin the web page associated with each published content,(iii) name of a text file that is distributed with the ac-tual content and is displayed by the BitTorrent softwarewhen opening the .torrent file. Our investigation indi-cates that the second approach (using the textbox) isthe most common technique among the publishers.Publisher’s Username: We browsed the Internet tolearn more information about the username associatedwith each top publisher. First, the username is in somecases directly related to the URL (e.g. user UltraTor-rents whose URL is www.ultratorrents.com). Thisexercise also reveals whether this username publisheson other major BitTorrent portals in addition to thePirate Bay. Finally, posted information in various fo-rums could reveal (among other things) the promotingweb site.Business Services: We characterize the type of ser-vices o"ered at the promoting URL and ways that theweb site may generate incomes (e.g. posting ads). Wealso capture the exchanged HTTP headers between aweb browser and the promoting URL to identify any es-tablished connection to third-party web sites (e.g. redi-rection to ads web sites or some third party aggregator)using the technique described in [14].

5.1 Classifying PublishersUsing the above methodology, we examined a few

published torrents for each one of the top publishersas well as sample torrents for 100 randomly selectedpublishers that are not in the top-100, called regularpublishers. On the one hand, we did not discover anyinteresting or unusual behavior in torrents published byregular publishers and thus conclude that they behavein an altruistic manner. On the other hand, a largefraction of seeded torrents by the top publishers system-atically promote one or more web sites with financialincentives. Our examination revealed that these pub-lishers often include a promotional URL in the textboxof the content web page. We classify these top publish-ers into the following three groups based on their typeof business (using the content of their promoting websites) and describe how they leverage BitTorrent portalsto intercept and redirect users to their web sites.Private BitTorrent Trackers: A subset of majorpublishers (25% of top) own their BitTorrent portalsthat are in some cases associated with private trackers[11]. These private trackers o"er a better user expe-rience in terms of download rate (compared to majoropen BitTorrent portals) but require clients to maintaincertain seeding ratio. More specifically, each participat-ing BitTorrent client is required to seed content propor-

tionally to the amount of data it downloads across mul-tiple torrents. To achieve this goal, users are required toregister in the web site and login before downloading thetorrent files. The publishers in this class publish 18% ofall the content while they are responsible for 29% of thedownloads. 2/3 of these publishers advertise the URLin the textbox at the content web page. Furthermore,they appear to gain financial profit in three di"erentways: (i) posting advertisement in their web sites, (ii)seeking donations from visitors to continue their basicservice, and (iii) collecting a fee for VIP access thatallows the client to download any content without re-quiring any kind of seeding ratio. These publishers typ-ically inject video, audio and application content intoBitTorrent portals. Interestingly, a significant fractionof publishers in this class (40%) publish content in a spe-cific language (Italian, Dutch, Spanish or Swedish) andspecifically a 66% of this group are dedicated to pub-lishing Spanish content. This finding is consistent withprior reports on the high level of copyright infringementin Spain [5].Promoting Web Sites: Another class of top publish-ers (23% of top) promote some URLs that are associatedwith hosting images web sites (e.g. www.pixsor.com),forums or even religious groups (e.g. lightmiddleway.com). These publishers inject 8% of the content and areresponsible of 11% of the downloads. Most publishers inthis class advertise their URL using the textbox in thecontent web page. Furthermore, most of these publish-ers (70%), specifically those that are running a hostingimage web site, publish only porn content. Inspectionof the associated hosting image web sites revealed thatthey store adult pictures. Therefore, by publishing porncontent in major BitTorrent portals, they are targetinga particular demography of users who are likely to beinterested in their web sites. The incomes of the portalswithin this class is based on advertisement.Altruistic Publisher: The remaining top publishers(52% of top) appear to be altruistic users since theydo not seem to directly promote any URL. These pub-lishers are responsible of 11.5% of the content and thesame fraction of downloads. Many of these users pub-lish small music and e-book files that require smalleramount of seeding resources. Furthermore, they typi-cally include a very extensive description of the contentand often ask other users to help with seeding the con-tent. These evidences suggest that these publishers mayhave limited resources and thus they need the help ofothers to sustain the distribution of their content.

In summary, roughly half of the top publishers adver-tise a web portal in their published torrents. It seemsthat their intention is to attract a large number of usersto their web sites. The incomes of these portals comefrom ads and in the specific case of private BitTorrentportals also from donations and VIP fees. Overall, these

Lifetime Avg. Publishing Rate(days) (torrents per day)

Private Portals 63/466/1816 0.57/11.43/79.91Promoting Web sites 50/459/1989 0.38/4.31/18.98Altruistic 10/376/1899 0.10/3.80/23.67

Table 4: Lifetime and Avg. Publishing Ratefor the di!erent classes of content publishers:BitTorrent Portals, Promoting Web Sites andAltruistic Publishers. The represented valuesare min/avg/max per class.

profit-driven publishers generate 26% of the content and40% of the downloads. Therefore, the removal of thissmall fraction of publishers is likely to have a dramaticimpact on the BitTorrent open ecosystem. Finally, afraction of publishers appear to be altruistic and re-sponsible for a notable fraction of published content anddownloads (11.5%). This suggests that there are someseemingly ordinary users who dedicate their resourcesto share content with a large number of peers in spiteof the potential legal implications of such activity.

5.2 Longitudinal View of Major PublishersSo far we focused on the contribution of major pub-

lishers only during our measurement intervals. Havingidentified the top publishers in our pb10 dataset, weexamine the longitudinal view of the contribution bymajor publishers since they appeared on the Pirate Bayportal. Toward this end, for each top publisher, we ob-tain the username page on the Pirate Bay portal thatmaintains the information about all the published con-tent and its published time by the corresponding usertill our measurement date (June 4, 2010)6. Using thisinformation for all top publishers, we capture their pub-lishing pattern over time with the following parameters:(i) Publisher Lifetime which represents the number ofdays between the first and the last appearance of thepublisher in the Pirate Bay portal, (ii) Average Publish-ing Rate that indicates the average number of publishedcontent per day during their lifetime.

Table 4 shows the min/avg/max value of these met-rics for the di"erent classes of publishers: Private Por-tals, Promoting Web Sites and Altruistic publishers.The profit-driven publishers (i.e. private portals andpromoting web sites) have been publishing content for15 months on average (at the time of the measurement)while the most longed-lived ones have been feeding con-tent for more than 5 years ago. Furthermore, some ofthese publishers exhibit a surprisingly high average rateof publishing content (80 files per day). The altruisticpublishers present a shorter lifetime and a lower pub-lishing rate that seems to be due to their weaker incen-6Note that we cannot collect information about fake pub-lishers since the web pages of their associated usernamesare removed by the Pirate Bay just after identifying theyare publishing fake content.

tives and their lower availability of resources.In summary, the lifetime of major publishers suggests

that content publishing in BitTorrent seems have been aprofitable business for (at least) a couple of years. Fur-thermore, the high seeding activity by profit-driven pub-lishers (e.g. private portals) over a long period of timeimplies a high and continuous investment for requiredresources that should be compensated by di!erent typesof incomes (e.g. ads) for these portals. We examinethe incomes of the profit-driven publishers in the nextsubsection.

5.3 Estimating Publishers’ IncomeThe evidences we presented in previous subsections

suggest that the goal of half of the top publishers is toattract users to their own web sites. We also showedthat most of these publishers seem to generate incomeat least by posting ads in their web sites. In essence,these publishers have a clear financial incentive to pub-lish content. In order to validate this key point, we as-sess their ability to generate income by estimating threeimportant but related properties of their promoting websites: (i) average value of the web site, (ii) average dailyincome of the web site, and (iii) average daily visits tothe web site. We obtain this information from severalweb sites that monitor and report these statistics forother sites on the Web. To reduce any potential errorin the provided statistics by individual monitoring websites, for each publisher’s web site we collect this infor-mation from six independent monitoring sites and usethe average value of these statistics across these webs7.

Table 5 presents the min/median/avg/max value ofthe described metrics for each class of profit-driven pub-lisher classes (i.e. private portals and promoting websites). The median values suggest that the promotedweb sites are fairly profitable since they value tens ofthousand dollars with daily incomes of a few hundreddollars and tens of thousand visits per day. Further-more, few publishers (<10) are associated to very prof-itable web sites valued in hundreds of thousand to mil-lions of dollars, that receive daily incomes of thousandsof dollars and hundreds of thousand visits per day. Insummary, these statistics confirm that these web sitesare indeed valuable and visible, and generate a substan-tial level of incomes.

6. OTHER BENEFICIARIES IN THE BIT-TORRENT MARKETPLACE

In previous sections we analyzed the main character-istics of major content publishers in BitTorrent, demon-strating that content publishing is a profitable busi-ness for an important fraction of the top publishers;7www.sitelogr.com, www.cwire.com, www.websiteoutlook.com,www.sitevaluecalculator.com, www.mywebsiteworth.com, www.yourwebsitevalue.com

Web site Web site Web siteValue ($) Daily Income ($) Daily visits

PrivatePortals 1K/33K/313K/2.8M 1/55/440/3.7K 74/21K/174K/1.4MPromotingWeb Sites 24/22K/142K/1.8M 1/51/205/1.9K 7/22K/73.5K/772K

Table 5: Publisher’s web site value ($), daily in-come ($) and num of daily visits for the di!erentclasses of profit-driven content publishers: Bit-Torrent Portals and Promoting Web Sites. Therepresented values are min/median/avg/maxper class.

a business that is responsible of 40% of the downloads.However, although content publishers are the key play-ers, there are other players who help sustain the busi-ness and obtain financial benefits including: Major Bit-Torrent Portals, Hosting Providers and ad companies.Figure 5 depicts the interactions between di"erent play-ers in BitTorrent content publishing where the arrowsindicate the flow of money between them. Next, webriefly describe the role of each player.Major Public BitTorrent Portals such as the Pi-rate Bay are dedicated to index torrent files. They arerendezvous points where content publishers and clientspublish and retrieve torrent files respectively. The mainadvantage of these major portals is that they o"er a re-liable service (e.g. they rapidly react to remove fakeor infected content). All this makes that millions ofBitTorrent users utilize these portals every day. Theseportals are the perfect target for profit-driven publish-ers in order to publish their torrents and advertise theirweb sites (potentially) to millions of users. Therefore,these major portals are one of the key players of the Bit-Torrent Ecosystem [22] that brings substantial financialprofit. For instance, the Pirate Bay is one of the mostpopular sites in the whole Internet (ranked the 94th inthe Alexa Ranking) as well as one of the most valuedones (around $10M).Hosting Providers are companies dedicated to rent-ing servers. Heavy seeding activity performed by somepublishers requires significant resources (e.g. bandwidthand storage). Thus a large fraction of major publish-ers obtain these resources from rented servers in hostingproviders who receive an income in return for the o"eredservice. Let’s focus on OVH, the ISP responsible of amajor portion of the content published in BitTorrent.Our measurement study shows that OVH contributesbetween 78 and 164 di"erent servers (i.e. IP addresses)across the di"erent datasets. Considering the cost of theaverage server o"ered by OVH in its web page (around300 e/month) we estimate that the average income ob-tained by OVH due to BitTorrent content publishingranges between roughly 23.4K and 42.9K e/month. Itis worth noting that some hosting providers have de-fined strict policies against sharing copyrighted mate-rial through P2P applications using their servers due

Figure 5: Business Model of Content Publishingin Bittorrent.

to legal issues8. However the income obtained by somehosting providers such as OVH seems to justify the riskof potential legal actions taken against them.Ad Companies are responsible for advertisements inthe Internet. They have a set of customers who wishto be advertised in the Internet and a set of web siteswhere they put their customers ads. They apply com-plex algorithms to dynamically select in which web siteand when to post each ad. They charge their customersfor this service and part of this income is forwarded tothe web sites where the ads have been posted. There-fore, ad companies look for popular web sites where toput ads for their costumers. We have demonstrated inthis paper that profit-driven BitTorrent content pub-lishers’ web sites are popular, thus most of them postads from ad companies9. Hence, part of the income ofad companies is directly linked to the BitTorrent con-tent publishing. Unfortunately, there is no practicalway to estimate the value of this income.

7. SOFTWARE FOR CONTENT PUBLISH-INGMONITORING

Using the methodology described in Section 2, wehave implemented a system that continuously monitorsthe Pirate Bay portal, and leverages RSS feed to quicklydetect a newly published content. Once a new torrent isdetected, our system retrieves the following informationfor that torrent: filename, content category and subcat-egory (based on the Pirate Bay categories), publisher’susername, and (in those cases we can) the publisher’sIP address as well as the ISP, City and Country associ-ated to this IP address. Furthermore, for profit-drivenpublishers described in this paper, we have created anindividual publisher’s web page that provides specificinformation such as the publisher’s promoted URL or

8http://www.serverintellect.com/terms/aup.aspx.9We have validated this by looking at the header exchangebetween the browser and the publishers’ web site servers.

business type. The system stores all this information ina database. Finally, we have built a simple web-basedinterface to query the resulting database. This interfaceis publicly available10 and access is granted to interestedparties by contacting the authors.

Our application has two goals. On the one hand, wewant to share this data with the research community topermit further analysis of di"erent aspects of the Bit-Torrent content publishing activity. On the other hand,we believe that this application can be useful for regularBitTorrent clients. First, a BitTorrent client can easilyidentify those publishers that publish content alignedwith her interest (e.g. an e-books consumer could findpublishers responsible for publishing large numbers ofe-books). Furthermore, we are working on implement-ing a feature to filter out fake publishers, allowing Bit-Torrent users of our application (in the future) to avoiddownloading fake content.

8. RELATED WORKSignificant research e"ort has focused on understand-

ing di"erent aspects of BitTorrent by gathering datafrom live swarms. Most of these studies have primar-ily examined demographics of users [12, 20, 22] andtechnical aspects of swarming mechanism [18, 13, 9].However, to our knowledge the socio-economics aspectsof BitTorrent that we addressed in this paper have re-ceived little attention. The most relevant work to thispaper is a recent study that examined the weaknessof BitTorrent privacy [6]. The authors analyzed thedemography of BitTorrent content publishers and pre-sented a highly skewed distribution of published con-tent among them as well as the presence of a signifi-cant fraction of publishers located at hosting providers.This indeed validates some of our initial observations.In another study, Zhang et al. [22] presented the mostextensive characterization of the BitTorrent ecosystem.This study briefly examined the demography of contentpublishers and showed a skewed distribution of the con-tributed content among them. The authors identify thepublishers by their usernames. We have shown that thisassumption may miss an important group of publisherswho post fake content, i.e. fake publishers. Our workgoes beyond the simple examination of demographics ofcontent publishers. We identify, characterize and clas-sify the major publishers and more interestingly revealtheir incentives and their motivating business model.

9. CONCLUSIONIn this paper we studied the content publishing ac-

tivity in BitTorrent from a socio-economic perspective.The results reveal that a small fraction of publishers areresponsible for 67% of the published content and 75% of

10http://bittorrentcontentpublishers.netcom.it.uc3m.es/

the downloads. We have carefully examined the incen-tives of major publishers and identified the following keycharacteristics: first, antipiracy agencies and malicioususers perform a systematic poisoning index attack overmajor BitTorrent portals in order to obstruct downloadof copyrighted content and to spread malware, respec-tively. Overall, this attack contributes 30% of the con-tent and attracts 25% (several millions) of downloads.Second, 37% of the (non-fake) published content is pub-lished by a small fraction of users that serve 54% ofthe (non-fake files) downloads. Our evidence suggeststhat these publishers have financial incentives for post-ing these contents on BitTorrent portals. The removalof these financial-driven publishers (e.g. by antipiracyactions) may significantly a"ect the popularity of theseportals as well as the whole BitTorrent ecosystem. Ifthis happens, will BitTorrent survive as the most popu-lar file-sharing application without these publishers?.

10. ACKNOWLEDGEMENTSThe authors would like to thank anonymous CoNEXT

reviewers for their valuable feedback as well as JonCrowcroft for his insightful comments. We are alsograteful for David Rubio’s help on developing the web-based application. This work has been partially sup-ported by the European Union through the FP7 TRENDProject (257740), the Spanish Government through theT2C2 (TIN2008-06739-C04-01) and CONPARTE (TEC2007-67966-C03-03) projects, the Regional Governmentof Madrid through the MEDIANET project (S-2009/TIC-1468) and the National Science Foundation under GrantNo. 0917381.

11. REFERENCES[1] http://www.theregister.co.uk/2008/06/02/

onk_further_arrests/.[2] http://yro.slashdot.org/yro/05/01/13/

2248252.shtml?id=123&tid=97&tid=95&tid=1.[3] MaxMind. http://www.maxmind.com.[4] The Impact of P2P File Sharing, Voice over IP,

Skype, Joost, Instant Messaging, One-ClickHosting and Media Streaming such as YouTubeon the Internet, 2007.http://www.ipoque.com/resources/internet-studies/internet-study-2007.

[5] Bay TSP Annual Report, 2008.http://tech.mit.edu/V129/N28/piracy/BayTSP2008report.pdf.

[6] S. Le Blond, A. Legout, F. Lefessant,W. Dabbous, and M. Ali Kaafar. Spying theworld from your laptop. LEET’10, 2010.

[7] D. R. Cho"nes, J. Duch, D. Malmgren,R. Guimera, F. E. Bustamante, and L. Amaral.Strange bedfellows: Communities in bittorrent.IPTPS’10, 2010.

[8] R. Cuevas, M. Kryczka, A. Cuevas, S. Kaune,C. Guerrero, and R. Rejaie. Is content publishingin bittorrent altruistic or profit-driven?. Techreport: http://arxiv.org/abs/1007.2327, 2010.

[9] R. Cuevas, N. Laoutaris, X. Yang, G. Siganos,and P. Rodriguez. Deep Diving into BitTorrentLocality. In ACM SIGMETRICS’10 (PosterSession).

[10] L. Guo, S. Chen, Z. Xiao, E. Tan, X. Ding, andX. Zhang. Measurements, Analysis, and Modelingof BitTorrent-like Systems. In ACM IMC’05.

[11] David Hales, Rameez Rahman, Boxun Zhang,Michel Meulpolder, and Johan Pouwelse.Bittorrent or bitcrunch: Evidence of a creditsqueeze in bittorrent? In WETICE ’09, 2009.

[12] M. Izal, G. Urvoy-Keller, E.W. Biersack, P.A.Felber, A. Al Hamra, and L. Garces-Erice.Dissecting bittorrent: Five months in a torrent’slifetime. In PAM ’04.

[13] S. Kaune, R. Cuevas, G. Tyson, A. Mauthe,C. Guerrero, and R. Steinmetz. UnravelingBitTorrent’s File Unavailability: Measurements,Analysis and Solution Exploration. In IEEEP2P’10.

[14] B. Krishnamurthy and C. E. Wills. On theleakage of personally identifiable information viaonline social networks. In ACM WOSN ’09, 2009.

[15] N. Laoutaris, D. Carra, and P. Michiardi. Uplinkallocation beyond choke/unchoke or how to divideand conquer best. In ACM CoNEXT’08.

[16] A. Legout, N. Liogkas, E. Kohler, and L. Zhang.Clustering and sharing incentives in bittorrentsystems. In ACM SIGMETRICS ’07.

[17] J. Liang, N. Naoumov, and K.W. Ross. The indexpoisoning attack in p2p file sharing systems. InIEEE INFOCOM’06, 2006.

[18] D. Menasche, A. Rocha, B. Li, D. Towsley, andA. Venkataramani. Content availability inswarming systems: Models, measurements andbundling implications. In ACM CoNEXT’09.

[19] M. Piatek, T. Isdal, T. Anderson,A. Krishnamurthy, and A. Venkataramani. Doincentives build robustness in BitTorrent? InNSDI’07, 2007.

[20] J.A. Pouwelse, P. Garbacki, D.H.J. Epema, andH.J. Sips. The BitTorrent P2P file-sharingsystem: Measurements and analysis. In IPTPS’05.

[21] M. Sirivianos, J. Han, P. Rex, and C.X. Yang.Free-riding in bittorrent networks with the largeview exploit. In IPTPS 07, 2007.

[22] C. Zhang, P. Dhungel, D. Wu, and K.W. Ross.Unraveling the bittorrent ecosystem. IEEETransactions on Parallel and Distributed Systems,2010.