Shedding Light on the Dark Corners of the Internet:
A Survey of Tor Research
, Junaid Qadirb, Muhammad U. Ilyasa,c
aSchool of Electrical Engineering and Computer Science (SEECS),
National University of Sciences and Technology (NUST), Islamabad-44000, Pakistan
bInformation Technology University, Lahore, Pakistan
cDepartment of Computer Science, Faculty of Computing and Information Technology,
University of Jeddah, Jeddah, Mecca Province 21589, Saudi Arabia
Anonymity services have seen high growth rates with increased usage in the past few years. Among various services,
Tor is one of the most popular peer-to-peer anonymizing service. In this survey paper, we summarize, analyze, classify
and quantify 26 years of research on the Tor network. Our research shows that ‘security’ and ‘anonymity’ are the most
frequent keywords associated with Tor research studies. Quantitative analysis shows that the majority of research studies
on Tor focus on ‘deanonymization’ the design of a breaching strategy. The second most frequent topic is analysis of path
selection algorithms to select more resilient paths. Analysis shows that the majority of experimental studies derived
their results by deploying private testbeds while others performed simulations by developing custom simulators. No
consistent parameters have been used for Tor performance analysis. The majority of authors performed throughput and
Keywords: Tor; Security; Anonymity; Survey; Analysis; Deanonymization; Breaching; Path selection; Performance
The Internet has revolutionized the world by trans-
forming it into a global entity. Widespread advantages of
Internet have spawned new industries and services. How-
ever, this connectivity comes at the cost of privacy. Every
Internet client has a unique identity in the form of an In-
ternet protocol (IP) address which can be translated to
its location by the local Internet service provider (ISP).
This lack of privacy has serious implications, particularly
for journalists, freedom ﬁghters and ordinary citizens.
Lack of privacy has lead to the use of anonymous com-
munication networks (ACN). ACNs hide client IP addresses
through various techniques. There are a number of ACNs
including Tor, Java anonymous proxy (JAP), Hotspot Shield
and UltraSurf etc. Among various ACNs, Tor is the one of
the most popular network, owing to its distributed nature
which makes it diﬃcult to connect the two end points of
a session. Recently, Tor has been used for bomb hoax at
Harvard . Similarly, it has been used by the Russians
to bypass online censorship . A number of attempts are
being made by FBI and other organizations to breach Tor
Email addresses: firstname.lastname@example.org (Saad Saleh),
email@example.com (Junaid Qadir),
firstname.lastname@example.org, email@example.com, firstname.lastname@example.org
(Muhammad U. Ilyas)
In this paper, we survey various studies conducted on
the Tor network covering the scope of these studies. We
quantify the studies into three broad but distinct groups,
including (1) deanonymization, (2) path selection, (3) anal-
ysis and performance improvements, and several sub-categories.
To the authors’ best knowledge, this is the most compre-
hensive attempt at analyzing Tor network research with a
focus over its anonymity mechanism. Table 1 presents a
comparison of this survey with previous surveys covering
the scope of researches and implementation (experiments),
veriﬁcation (simulations) and analysis of various research
works. Categorization of ﬁrst column is made by listing
all Tor areas considered in our study. AlSabah and Gold-
berg  presented the most comprehensive study covering
complete Tor network and our paper is complementary to
their survey paper. However, our paper pays more focus
to the anonymity and breaching aspects of Tor than their
paper. Their research paper presents only twenty refer-
ences related to anonymity while we present more than
Analysis of keywords used in various studies shows
that anonymity, security and privacy have been used the
most. Our study shows that majority of the research
works have been made in the ﬁeld of “deanonymization”
1This paper has 146 references, some of the references are to tools
rather than research works; in all, we are considering a research
corpus of 120 references.
Preprint submitted to Journal of Network and Computer Applications March 16, 2018
Table 1: Comparison of other surveys with this survey. Tick mark indicates coverage while blank space indicates non-coverage of topics in
various research works. Coverage (Studies) indicate number of research works referred in any survey.
Research, Year, Coverage (Studies)
2017, 2016, 2016, 2015, 2014, 2014, 2010, 2007,
120 120 10 99 40 37 109 32
Hidden services X X X X X
Finger printing X X X
Attacks X X X X X X X
Traﬃc analysis X X X X X X X
Improvements X X X
Bypassing Tor X X X X X X
Path Selection Algorithm design X X X
Path analysis X X X
General X X X X X X
Evaluation X X
Improvement X X
Mobile Tor X
Parameters X X
Private Setup X
Cloud services X
UC framework X
Custom Simulator X
Shadow simulator X
track, followed by “performance analysis and architectural
improvements”. In the deanonymization track, a major
chunk of research is devoted to breaching attacks followed
by traﬃc analysis. In the path selection track, most re-
search works focused on the development of new algo-
rithms. Relays, protocol messages and traﬃc intercep-
tion have been the most frequently exploited factors in the
Tor’s deanonymization track. In the path selection track,
performance and anonymity have been the most commonly
used factors. Performance, relay selection and anonymity
have been the most studied parameters in the performance
analysis and improvement track. Analysis over simulations
and experiments shows that 60% of studies used experi-
ments and 86% of those experiments were carried out on
private testbed networks. Among simulations, 75% of the
studies developed their own simulator to analyze Tor net-
work. Analysis of simulation parameters shows that there
is no distinct pattern of parameters. However, majority of
the studies used bandwidth and latency.
Table 2 presents a glossary of the important abbrevia-
tions used in our survey paper. This paper is organized as
follows: Section 2 introduces the architecture of Tor net-
work and its comparison with other anonymity services.
Section 3 presents the studies covering deanonymization,
path selection, and performance analysis and architectural
improvements. Section 4 presents the simulations and ex-
periments conducted in previous studies. Section 5 presents
the Tor performance metrics, our ﬁndings and open re-
search areas in Tor. Finally, section 6 concludes the paper.
2. Architecture and comparison of Tor with other
In this section, we present and discuss Tor and other
anonymity tools. In the ﬁrst part, we present the archi-
tecture of Tor network before presenting details of the re-
search in Tor. In the second part, we present the com-
parison and working mechanism of other anonymity tools
which compete with Tor.
2.1. Architecture of Tor network
Tor network is composed of a decentralized distributed
network of relays operated by volunteer users . In July
2016, nearly 10,000 users (per day) participated in the
Tor network (as Tor relays and Tor bridges) to provide
anonymity services to nearly half a million users daily .
History of Tor dates back to late 1990’s when Goldschlag,
Reed and Syverson presented the architecture and imple-
mented onion routing in several papers [15, 16, 17, 18]
which laid the foundation of Tor network by providing
Table 2: Glossary of the important abbreviations used in the text.
ACK Acknowledgement MRA Multi-Resolution Analysis
ACN Anonymous Communication Network NAT Network address translation
ADSL Asymmetric digital subscriber line NTP Network Time Protocol
AES Advanced Encryption Standard OP Onion Proxy
AS Autonomous System OR Onion Router
CGI Computer-generated imagery PGP Pretty Good Privacy
CSRF Cross site request forgery POP3 Post Oﬃce Protocol 3
DHCP Dynamic Host Conﬁguration Protocol PPTP Point-to-Point Tunneling Protocol
DNS Domain Name System P2P Peer-to-Peer
DoS Denial of Service QoE Quality of Experience
DS Directory Server QoS Quality of Service
DSL Digital Subscriber Line ROC Region of Convergence
EWMA Exponentially weighted moving average RRD Round Robin Database
FIFO First In First Out RTT Round Trip Time
FN False Negative SMTP Simple Mail Transfer Protocol
FP False Positive SSH Secure Shell
HTML HyperText Markup Language SVM Support Vector Machine
HTTP Hypertext Transfer Protocol TAP Tor Authentication Protocol
IMAP Internet Message Access Protocol TCP Transmission Control Protocol
ICMP Internet Control Message Protocol TMT Tunable mechanism of Tor
IP Internet Protocol Tor The Onion Router
ISP Internet Service Provider TP True Positive
I2P Invisible Internet Project TTL Time To Live
JAP Java Anonymous Proxy URL Uniform Resource Locator
JVM Java virtual machine VDE Virtual Distributed Ethernet
LAN Local area network VM Virtual Machine
L2TP Layer 2 Tunneling Protocol VPN Virtual Private Network
ML Machine Learning VPS Virtual Private Server
proxy servers which are resilient to eavesdropping and ef-
fectively hide client’s IP address.
The Tor network consists of routers which cooperate
with each other to provide low latency anonymity services
to users. Central servers help Tor establish and update
links between Tor routers. User participation as Tor re-
lays (router) is optional, but it is recommended because
it improves the chances of staying anonymous, because it
increases the traﬃc to the user.
Tor’s architecture has three types of components, namely
onion proxy (OP), onion router (OR) and directory server
(DS). OPs are used by Tor users to obtain up-to-date in-
formation of operating relays from DS. OPs also creates
connections using the information from a DS. Users may
conﬁgure OPs to select speciﬁc routers.
ORs are Tor relay routers, operated by volunteer users,
to act as entry (guard), middle and exit relays. Informa-
tion of all online relays is available at DS. To counter at-
tacks on Tor that block Tor relays, a secret group of Tor
relays exists with the DS, called bridges. A set of three
bridge relays is available through unique Gmail addresses.
Once a connection is established, every OR knows only
immediate predecessor and successor node.
Nine authorities acting as Tor DSs keep an up-to-date
record of all available ORs and broadcast the bandwidth,
IP, public key, exit policies etc. to OPs.
Entry RelayClient (OP)
Figure 1: Architecture of Tor network.
Figure 1 shows a circuit established from a user to a
server through three Tor relays. The various steps involved
in circuit establishment are listed below:
1. OP sends HTTP requests to DS for information about
2. OP selects three Tor routers (entry relay, intermedi-
ate relay and exit relay) using Tor’s path selection
algorithm considering maximum anonymity and per-
3. OP sends a Create Cell request (containing half of
Diﬃe-Hellman handshake ) to the entry node.
The entry node replies with the hash of the negoti-
4. Next, OP sends a Extend Cell request to the entry
node containing the address of the intermediate re-
lay and encryption key. Entry node forwards the
cell to the intermediate node. Similar to previous
case, the intermediate node replies with Created cell
response. Similar process continues till client nego-
tiates the key with the exit relay.
5. OP constructs IP packet P1 with source and des-
tination IP addresses of exit relay and destination
server, respectively, and packet size of 512 bytes.
6. OP encrypts the packet further with the key E3, ne-
gotiated between client and exit relay, and contain-
ing source and destination addresses of intermediate
relay and exit relay.
7. Next, OP encrypts with the key E2, negotiated be-
tween client and intermediate relay, and containing
source and destination IPs of entry relay and inter-
8. Finally, OP constructs packet P2 encrypted with key
E1, negotiated between client and entry relay, and
containing source and destination IPs of OP and en-
9. Packet P2 is transmitted from the entry relay, which
decrypts the packet and forwards it to intermediate
relay. All relays decrypt the packet using their spe-
cialized decryption keys and forward it towards the
Cell refers to the Tor Packet, comprising of payload
data and headers, with an aggregate size of 512 bytes [20,
21, 22]. Padding is used to ﬁll cells with less data.
Tor relays communicate with each other by pairwise
TCP connections. Traﬃc multiplexing is used to trans-
fer data between any pair of relays. Tor employs token
buckets to rate limit connections. Buckets are ﬁlled and
removed with tokens based on conﬁgured bandwidth limits
and data read, respectively. TCP buﬀers are read using
a round-robin scheduling mechanism. For ﬂow control,
edges (client and exit node) keep track of data ﬂow by
maintaining an active window about the packets in ﬂight.
Data packets are processed in a ﬁrst-in-ﬁrst-out (FIFO)
manner from the queues of Tor relays. Multiplexing of
packets, from Tor relays to relay links, is performed using
exponentially-weighted moving average (EWMA) sched-
2.2. Comparison with other anonymization services
In this section, we present the features and working
mechanisms of other deanonymization services which com-
pete with Tor network. Table 3 presents a comparison of
Tor with other deanonymization services. Table 3 shows
that Tor is the only anonymity service which provides var-
ious services (http, https, visible TCP port, remote DNS,
hides IP and user-to-proxy encryption) under all circum-
stances. On the contrary, JonDo, I2P, CGI and socks5
provide some services in limited circumstances only. A
summary of various anonymity services is presented in fol-
2.2.1. Cross Platform Anonymity Tools
A number of cross platforms anonymity tools are used
now-a-days. In below lines, we summarize the basic work-
ing mechanism of prominent anonymity tools.
•Java Anonymous Proxy (JAP or JonDonym) :
Users can select among several Mix Cascades, diﬀer-
ent from P2P.
•PacketiX.NET : Virtual LAN card and Virtual
HUB by Ethernet and can provide layer 2 VPN vir-
•JanusVM : Uses Tor for all TCP based con-
nections including DNS requests and provides web
•proXPN : A VPN provider which hides the IP
address of user and encrypts the data.
•USAIP : A VPN service provider with servers in
Switzerland, Luxembourg and Hungary etc.
•VPNReactor : Uses a VPN connection with time
limits for free and pro service and user logs are kept
for 5 days.
2.2.2. Windows Based Anonymity Tools
A number of windows based anonymity tools are large
competitors of Tor network. Basic mechanisms of promi-
nent anonymity tools are summarized in below lines:
•xB Browser : A browser designed to run over the
Tor network and XeroBank anonymity network.
•Hotspot Shield : Uses VPN. Hosts web servers
accessible through proxy and has a central server
that can be compromised.
•AdvTor : Acts as a portable client and server for
the Tor network. Improvements include the UNI-
CODE path, HTTP and HTTPS protocols, estimates
AS paths etc.
•SecurityKISS : A VPN service based on Open-
VPN, PPTP and L2TP.
•UltraSurf : Uses HTTP proxy to bypass censor-
ship and uses encryption protocols for privacy.
•CyberGhost VPN : OpenVPN based proprietary
client, Centralized server with VPN using 1024-SSL
•Freegate : Uses range of proxy servers (called
Dynaweb) along with encryption.
Figure 2: Word cloud of keywords used in Tor research works.
Table 3: Comparison of Tor with other anonymization services (‘X?’ refers to ‘In limited circumstances’).
HTTP HTTPS Visible
Hides IP user-to-
https X X X X?
socks4 X X X
socks4a X X X X
socks5 XX XXX
CGI X?X?X X?X?
I2P X?X?X X X X
JonDo X X X?X X X
Tor X X X X X X
2.2.3. Linux based solutions
Linux, being the prominent platform, is used by vari-
ous anonymity tools to guarentee anonymity to its users.
Working mechanism of some tools is summarized in below
•Tails (Amnesic Incognito Live System) : Has a
Debian Linux distribution using Tor network.
•Privatix : Provides encryption with anonymous
web browsing (using Tor, Torbutton and ﬁrefox).
2.2.4. Anonymous Search Engines
Several search engines also provided anonymity to their
users by collecting no user information. Some of these
anonymous search engines are as follows:
•Ixquick : Opens all search results through a proxy
•DuckDuckGo : Provides searcher’s privacy and
avoids “ﬁlter bubble”. Displays same information to
all users for a given search.
2.2.5. Anonymous Emails
Anonymous emails are provided by many servers. In
following lines, we have summarized the tools provided
anonymous email capability to their users.
•Anonymous E-mail : Sends the emails with anony-
•Safe-mail : Provides a secure communication,
storage, distribution and sharing system for the in-
•HushMail : Provides a PGP encrypted email ser-
•10 Minute Mail : Gives an email address which
expires after 10 minutes to counter spam emails.
•Yopmail : Provides a disposable email address
with no registration and password.
3. Tor Research Areas
A large number of networks have utilized Tor networks
for various purposes. To get an overview of the research
Anonymity without Tor
Path Selection 
Traffic Analysis [2,3,4]
Traffic Identification [5,6,15]
Relays Identification 
Bridges Identi fication 
Entry and Middle Relay [9,11]
Compromised Exit Re lay
Malicious Nodes, Apps., Servers and
Decoy Traffic [10,20, 21 ,24,26]
Destroy/DNS Requests 
DoS Attack Variations [17,19,23]
Autonomous S ystems 
Correlation Functions [27,20,28,29 30 34]
Machine Learning [30,31,32,88]
Signal Processing [ 33]
Relays and Exit Nodes [16,35,36]
Key Exchange Mechanism 
Transport Layer [3 8,39]
Counter DoS Attack [40,19]
In-network Anonymization 
Relays, Latency, Performance, Hops, Load,
Capacity, Guard Usage Time, Link
Bandwidth [43,44,49,50,54,46, 48,53,56 ]
User Demand for An onymity vs
Performance [47 ]
Bypassing Self-adve rtised Bandwid th
Path Compromise 
Relay Selection Mechanism 
Bridge/Relay Selection 
Architecture Design 
Usability Analysis [61,62,91 ]
Topology and Hosts 
Simulation Tools [70,71]
Delays, Bandwidt h, QoS
Exit Policies [76,79]
Node Selection 
Hidden Services [8 0,81]
Authentication Protocol 
Relay Count [83,84]
Bridges [85,86 ]
Anonymity [87,88 ]
Throughput and Latency [90,92,93]
Figure 3: Taxonomy of Tor research. Tor literature can be broadly classiﬁed into three areas: deanonymization, path selection, and perfor-
mance analysis and architectural improvements.
areas dealt in various studies, we created a word cloud of
the keywords of these studies. Figure 2 shows the word
cloud of keywords (on log-scale) of all studies cited in this
paper. Data of keywords shows that anonymity, privacy
and security are the most important terms dealt in various
In our review, we observed that research works on
Tor could be broadly classiﬁed into three tracks/categories
which include (1) deanonymization, (2) path selection, and
(3) Analysis and performance improvement. Figure 3 shows
the classiﬁcation of our survey paper along with a list of
all research works present in various subcategories.
In deanonymization track, research works were observed
in six diﬀerent categories covering (1) Hidden services which
limit their scope to hidden servers identiﬁcation, (2) Fin-
gerprinting which are based on pinpointing Tor network,
(3) Attacks which are focused over breaching Tor network,
(4) Traﬃc analysis which analyze Tor traﬃc to pinpoint
the weaknesses, (5) Studies studying improvements in Tor
to avoid deanonymization, and (6) Anonymity without Tor
which suggest alternate methods to provide anonymity by
pinpointing weaknesses in Tor.
In the path-selection track, all research works are either
based upon (1) Development of new algorithms providing
better eﬃciency and anonymity, and (2) Analysis of Tor’s
algorithm to study its strong and weak points in circuit
Lastly, analysis and performance improvement track
focuses on four sub-areas which include (1) Generalized
studies over Tor providing usability and social implica-
tions, (2) Modelling studies which focus on the develop-
ment of model for analysis of Tor, (3) Analysis studies
which cover QoS, relays, servers, etc., (4) Performance im-
provement studies provide modiﬁcation in relays and ar-
chitecture to provide better QoS, and (5) Development of
eﬃcient mechanisms for Tor clients with mobility.
Figure 4 shows the classiﬁcation of various research ar-
eas studied in the Tor network. It is pertinent to mention
that all numeric values used for all pie charts, ﬁgures and
tables in this paper have been calculated by the authors.
Source of all numeric values is the ‘Reference’ section at
the end of the paper, which includes scholarly research
articles. Moreover, references have been collected by the
authors from Tor repository2with a particular bias to-
wards papers covering ‘Tor’ network only. Span of col-
lection varies from 2007 to 2017 in reputed international
conferences and journals. We also included the important
studies in this ﬁeld before 2007 which play helpful role in
Figure 4: Classiﬁcation of Tor research areas.
understanding of Tor network, such as . Many ar-
ticles were also collected from ‘ACM digital library’ and
‘IEEE Xplore digital library’ with a particular focus to-
wards anonymity and security in Tor.
3.1. Tor Deanonymization
Breaching the Tor network is one of the most widely
studied research problems. In fact, the majority of the
studies describe deanonymization attacks without identi-
fying any counter-measures . Research works covering
deanonymization can be subdivided into a number of sub-
categories including (1) Hidden services identiﬁcation, (2)
Tor traﬃc identiﬁcation, (3) Attacking Tor network, (3)
Traﬃc ﬁngerprinting, (4) Focusing over Tor improvements,
and (6) Providing anonymity without Tor. Classiﬁcation
of various research problems is shown in the pie chart in
Table 4 presents a comparison of various research works
in the Tor’s deanonymization track. Prominent patterns
show that relay compromise and traﬃc interception are the
most frequent factors in deanonymizing Tor. This suggests
that relays and traﬃc are more susceptible for exploitation
than other factors. Individual details of various research
works in following subsections would explain this exploita-
tion in much detail.
3.1.1. Tor Hidden Services
An important feature of the Tor network is provision-
ing of Tor service through a hidden server. A series of
protocols used by hidden server and Tor users can make
location of hidden server invisible to client . However,
several studies listed below address the deanonymization
of hidden servers.
Locating Hidden Server: Overlier and Syverson 
presented new attack strategies to detect the location of
hidden servers using only one Tor node. They proposed
changes in route selection and relay selection to increase
anonymity. The average duration of the attack varied from
minutes to a few hours. The various attacks they consid-
ered included the timing signature analysis attack, service
location attack, predecessor attack and distance attack.
Their proposed solution included introducing middleman
nodes to connect to rendezvous points, introducing dummy
Figure 5: Classiﬁcation of Tor’s deanonymization approaches.
traﬃc, extending hidden server path to rendezvous point
and using guard entry nodes.
Timing Signature Attack: Elices et al.  presented a
ﬁngerprint analysis attack for Tor’s hidden services. They
used timestamps from logs of machines hosting hidden ser-
vices on the Tor network to generate detectable ﬁnger-
prints. The authors studied delay properties of the Tor
network and other users’ log entries to make the ﬁnger-
print attack feasible.
Application Layer Correlation Attack: Zhang et al.
 described an application level HTTP-based attack for
Tor’s hidden services. Time correlation was used to assess
the resemblance between web accesses and the traﬃc gen-
erated in a compromised Tor router. This attack assumes
that the compromised onion router can operate as an entry
Detection, Measurement and Deanonymization of hid-
den services: Biryukov et al.  analyzed weaknesses
in hidden services which can be exploited by attackers to
detect, measure and deanonymize hidden services running
over the Tor network. Services of three diﬀerent appli-
cations were analyzed, (1) Botnet for command and con-
trol, (2) Silk Road3and (3) DuckDuckGo4. The study
identiﬁed major ﬂaws of Tor that included the inﬂation
and cheating of bandwidth by a corrupted relay node, and
cheating marking mechanism of ﬂags in Tor network from
attacker relay node.
3.1.2. Tor Traﬃc Detection
A number of studies focus their research on the iden-
tiﬁcation of Tor traﬃc form other network traﬃc. These
studies suggest that diﬀerentiation of traﬃc can ultimately
be used to block Tor traﬃc, as done by China a number
of times in the recent past . Various approaches for
traﬃc identiﬁcation are summarized as follows.
Tor Traﬃc Identiﬁcation from Network Traﬃc: Bai
et al.  studied the traﬃc identiﬁcation mechanisms
3Silk road was an online market place which provided anonymity
to its customers by way of the Tor network. It was used in great
part for the sale of drugs and illegal materials and was shutdown by
the FBI. Defunct Website: http://silkroad6ownowfk.onion
4DuckDuckGo is a search engine that does not track its users and
provides anonymity to users by giving same search results for any
query to all users. Website: https://duckduckgo.com
Table 4: Research works focused on Tor’s deanonymization. Table entries symbolize attacks (Att.), counter-attacks (Cou. Att.), Analysis
(Ana.), relays (Rel.), Autonomous systems (AS), browser (brows.), server (serv.), decoy traﬃc (Dec. Traf.), protocol messages (Prot. Mess.),
traﬃc interception (Traf. Interc.), Flag cheating (Flag Cheat.).
Research Focus Exploited Tor’s weakness Idea
Att.Cou.Ana. Compromised Dec.Prot. Traf. Flag
Att. Rel.AS brows.serv.traf.mess.interc.Cheat.
Overlier and Syverson  X X X Timing signature analysis attack, service location attack,
predecessor attack and distance attack
Elices et al.  X X Fingerprint analysis attack
Zhang et al.  X X Application level time correlation attack
Biryukov et al.  X X X Using corrupted relay node and cheating Tor’s ﬂag marking
Tor’s Traﬃc identiﬁcation
Bai et al.  X X Packet examination, context checking and matching
Barker et al.  X X Using unsupervised machine learning techniques over packet sizes
AlSabah et al.  X X Application level time correlation attack
Houmansadr et al.  X X Passive and active attacks to bypass traﬃc imitation technique
Chakravarty et al.  X X Observing bandwidth ﬂuctuations through compromised node
Winter and Lindskog  X X Using port tuples
Sulaiman and Zhioua  X X Using unpopular ports over compromised relays
Chan-Tin et al.  X X X Using malicious servers to observe traﬃc ﬂuctuations over relays
Pries et al.  X X X Passing duplicate cells through compromised entry relay
Wang et al.  X X Returning malicious page through compromised exit relay
Wagner et al.  X X X Compromised exit node to send images which is used by
semi-supervised learning algorithm
Benmeziane and Badache
X X Using destroy and DNS requests for man-in-the-middle attack
Jansen et al.  X X Using valid protocol messages over relays to perform denial of
Abbot et al.  X X X Use compromised relay and user’s browser for man-in-the-middle
Bauer et al.  X X Advertise low bandwidth to divert traﬃc towards malicous nodes
Edman and Syverson  X X Using autonomous systems to breach Tor traﬃc
Barbera et al.  X X Perform DoS attack by placing large load on Tor routers
Le Blond et al.  X X Using peer-to-peer applications with compromised exit relays to
Geddes et al.  X X Exploiting compromised exit node to advertise high stats to
attract large traﬃc
Chakravarty et al.  X X Using decoy traﬃc to detect traﬃc interception
Tor traﬃc analysis attacks
Johnson et al.  X X Correlation based attacks using a single compromised relay
Murdoch and Danezis  X X X Use timing signature attack by passing large traﬃc from
corrupted node to Tor relays
Chakravarty et al.  X X X Generate traﬃc between two servers and map relays through
Zhang et al.  X X Suggested priority queue algorithm to bypass correlation
between load and latency
Song et al.  X X Use time and stream size with k-means algorithm to
Panchenko et al.  X X Use volume, time and direction for classiﬁcation
Wang and Goldberg  X X Use Tor cells for website ﬁngerprinting
Jin and Wang  X X use wavelet based decomposition to estimate timing distortion
Gilad and Herzberg  X X Breach by oﬀ-path TCP connection or eavesdrop on clients
Gros et al.  X X X Proposed Honeywall to rank node’s reliability
Winter and Lindskog  X X Proposed exit relay scanner to avoid misuse of exit node
Xin and Neng  X X Proposed a tuning mechanism to keep a track of reliable nodes
Backes et al.  X X X Identiﬁed ﬂaws in current key exchange mechanisms
Marks et al.  X X Suggested separate bi-directional TCP links to increase
Nowlan et al.  X X X Suggested use of uTCP and uTLS to avoid head of line blocking
Danner et al.  X X X Investigated DoS attack and proposed improvements to avoid it
Anonymity without Tor
Herzberg et al.  X X X Suggested camouﬂaged web server by mimicking GMAIL traﬃc
Mendonca et al.  X X Proposed concealed source identiﬁer through network service
of popular anonymity tools, i.e., Tor and Web-Mix. Au-
thors used ﬁngerprint identiﬁcation (packet examination
and packet context checking) followed by matching to iden-
tify the traﬃc. Key attributes used for traﬃc identiﬁcation
from other network traﬃc include speciﬁc strings, packet
length and packet transmission frequency in the network.
Diﬀerentiate Tor Traﬃc from Encrypted Traﬃc: Barker
et al.  showed that traﬃc from the Tor network can be
diﬀerentiated form encrypted traﬃc in the network. They
captured regular HTTPS, Tor HTTPS and HTTP traﬃc
routed through Tor and analyzed their packet sizes and de-
veloped an unsupervised machine learning (ML) classiﬁer
that operates only on packet size attribute with 97.54%
true positive (TP) and 1.06% false positive (FP) rates.
Diﬀerentiating Tor Traﬃc: AlSabah et al.  de-
veloped an ML classiﬁer to diﬀerentiate web traﬃc from
bulk download traﬃc. AlSabah et al. used the follow-
ing four features to classify Tor traﬃc: (1) Circuit life-
time, (2) data transferred, (3) cell inter-arrival times, and
(4) recently sent cells. They tested na¨ıve Bayes, Bayesian
networks and decision tree classiﬁers. Using the proposed
classiﬁcation method, they reported 75% improvement in
responsiveness and 86% decrease in download rates.
Fingerprinting Tor traﬃc: Houmansadr et al. 
aimed to diﬀerentiate the traﬃc of anonymous networks
from other network traﬃc. They claimed that mimicking
other traﬃc is an obsolete way for anonymity. They de-
vised a number of passive and active attack strategies to
breach anonymous networks. Their study suggested the
use of partial imitation and use of new strategies by incor-
porating popular protocols like HTTPS email etc.
Tor Proxy Node Identiﬁcation: Chakravarty et al. 
described a novel attack that identiﬁes all Tor relays par-
ticipating in a given circuit. The attack modulates the
bandwidth of an anonymous connection through a com-
promised server, router or an entry point and observes the
resultant ﬂuctuations in the Tor network using LinkWidth
. LinkWidth sends a train of pulses comprising of al-
ternate TCP-SYN and TCP-RST packets and capacity is
computed at the receiver end by estimating packet disper-
sion. Authors reported a 59.46% TP rate and 10% true
negatives rate for compromised Tor relays using the pro-
Identiﬁcation of Tor Bridges: Winter and Lindskog
 conducted an extensive investigation into the block-
ing of Tor relays and bridges by China. Their investiga-
tion showed that Tor bridges were blocked by port tuples,
rather than IP addresses and that bridges were blocked
only when they were active. Their investigation also showed
that adversaries did not conduct traﬃc ﬁngerprinting for
domestic traﬃc and that packet fragmentation could be
used to circumvent China’s ﬁrewall.
Fingerprinting Keywords in Search Queries: Oh et al.
 investigated the viability of keyword ﬁngerprinting at-
tacks in the Tor network. Study showed that eﬀective fea-
ture selection can help any passive adversary in ﬁguring
out the identity of the user. Time and volume of traﬃc
play the most crucial role in determining the identity of
the user. Among other features keyword sets, incremental
search and high security search are other features used for
classiﬁcation. For experiments, authors collected a key-
word dataset containing 160,000 search queries of google.
For one of 300 targeted keywords of Google, experimen-
tal results demonstrated recall, precision and accuracy of
80%, 91% and 48%, respectively.
3.1.3. Tor Attacks
Attacking the Tor network is an interesting research di-
mension which ultimately aims to block access to it. Sev-
eral attempts by China and other countries have failed in
the recent past because Tor is being improved continuously
. In this subsection, we summarize various studies cov-
ering Tor attacks.
Unpopular Ports Attack: Sulaiman and Zhioua  de-
scribed an attack they developed which can compromise
circuits in the Tor network. Their attack takes advan-
tage of unpopular ports in the Tor network. Sulaiman and
Zhioua added a small number of compromised entry /exit
relays to the Tor network (∼30 relays) which permit the
use of unpopular ports. By doing so, 50% of developed
circuits can be compromised, which signiﬁcantly decreases
the anonymity of the Tor network.
Circuit Clogging Attack: Chan-Tin et al.  proposed
an attack that can identify the Tor routers used in any
circuit. For the proposed attack a client connects to a
malicious server which sends data to the client in large
bursts and in small amounts. During large bursts, Tor
routers take long times to process the extra amount of
data. Authors showed that continuous monitoring of all
Tor relays can identify the Tor relays used in the particular
circuit. A mechanism to detect the behavior of malicious
routers by the client was also evaluated, which measured
network latency of the client.
Replay Attack: Pries et al.  suggested a replay at-
tack to detect the exit routers in the Tor network. The re-
play attack assumes that the entry onion router is compro-
mised. The replay attack duplicates packets coming from
a sender. Tor uses counter-mode Advanced Encryption
Standard (AES-CTR)5for encryption and decryption,
any duplicate cells will give a cell recognition error at the
exit routers. This behavior leaks exit router information
to the entry router by simple correlation.
Flow Multiplication Attack: Wang et al.  designed
a ﬂow multiplication attack similar to a man-in-the-middle
attack. The attack assumes that the exit router is com-
promised. Whenever a client sends a request to target
server, the exit router returns a malicious page which trig-
gers certain fetch requests in the client browser over the
same circuit. An accomplice at the entry router can see
5Advanced Encryption Standard (AES) is an encryption standard
which is based upon substitution-permutation technique. It has three
members Rijndael family each with block size of 128 bits and key
lengths of 128, 192 and 256 bits.
the requests, and together with knowledge of the exit relay,
identify the complete Tor circuit.
Attack Using Game Theory and Data Mining: Wag-
ner et al.  proposed an attack which exploits the exit
malicious exit node to cluster observed traﬃc ﬂows using
an active tag injection scheme. The proposed method has
two steps, (1) image tags are injected into HTML replies
from the exit node to the user, and (2) a semi-supervised
learning algorithm based upon deep data mining is used
to reconstruct the entire browsing session of the user. The
authors model the Tor network in form of a game theo-
retical concept where all Tor users and rogue nodes play
a game for identiﬁcation of malicious node. Once a rogue
node has been identiﬁed, it’s game is over because no other
user uses it due to presence of special ﬂag in it. Authors
main aim is to work over over the equilibrium between
rogue nodes and Tor users.
Attack using Destroy and DNS Requests: Benmeziane
and Badache  investigated possible breaches of Tor tar-
geting its network requests. They exploited destroy re-
quests (Tor’s circuit destruction requests) and DNS re-
quests to break anonymity. Destroy requests are not en-
crypted, which poses a serious threat to Tor. Moreover, a
local eavesdropper can use the man-in-the-middle attack
strategy against DNS requests, which are unprotected.
The Sniper Attack: Jansen et al.  presented the
Sniper attack, a low-resource denial-of-service (DoS) at-
tack against the Tor network which can disable arbitrary
relays. The adversary builds a Tor circuit through the tar-
get relay and starts obtaining a large ﬁle by continuously
sending the SENDME cells (protocol messages for contin-
uously receiving the ﬁle), which increases the congestion
window size. By repeating over multiple circuits, memory
of host of target relay would exhaust which can disrupt
the functioning of Tor relay. Experiments showed that an
adversary can consume 2,187 KB/s memory of a victim
relay at the cost of very little bandwidth and decrease Tor
network bandwidth by as much as 35%.
Browser-based Attacks: Abbot et al.  proposed a
novel attack that tricks a user’s web browser into sending
a distinctive signal over the Tor network (by installing a
Java or HTML script). An attacker that controls an exit
relay can use it in a man-in-the-middle attack to mirror
and forward duplicated traﬃc to a malicious server. By
analyzing the data, the malicious server can deanonymize
the Tor user. However, this study makes two signiﬁcant
assumptions: the ability to control the exit relay and the
ability to conﬁgure / compromise a targeted user’s web
Congestion Attack using Long Paths: Evans et al. 
proposed an extension to the congestion attack proposed
by Murdoch and Danezis  owing to the enormity of
the current Tor network. Evans et al. proposed the com-
into a user’s browser, which makes the browser send a re-
sponse every second. They suggested modiﬁcations such
ability to control latency of routers. In the modiﬁed de-
sign, routers keep a track of all paths with ﬂags and disable
any request for latency by using ﬂags.
Exploiting Routing Algorithm: Bauer et al.  ex-
ploited Tor’s routing algorithm to steer a disproportion-
ate number of users towards selecting their entry and exit
relays from a set of malicious Tor routers. Bauer at al.
suggested that low-latency constraints force Tor’s routing
algorithm to prefer nodes advertising high bandwidths. In-
stead of performing complex traﬃc analysis techniques,
the authors suggested to collect detailed ﬂow logs from
malicious nodes (both entry and exit nodes) and use the
information of node selection to deanonymize ﬂows.
Analyzing Autonomous Systems for Tor Path Selec-
tion: Edman and Syverson  analyzed the eﬀect of au-
tonomous systems (AS) for path selection in Tor network.
They studied the selection of AS residing in diﬀerent coun-
tries and found it quite eﬀective. Traﬃc analysis of the Tor
network showed that majority of traﬃc passes through a
few ASs because all established paths focus over latency
and anonymity which occurs better in some ASs. Anal-
ysis shows that increase in relays has not increased the
diversity to a large extent.
DoS Attack using Cell Flooding: Barbera et al. 
presented a novel attack which generates a few circuits re-
quiring large computing and networking resources. Their
study showed that this attack requires only 0.2% resources
for old routers and 1 −16% router resources for new at-
tacks, which makes it an inexpensive attack to execute.
Barbera et al. proposed a mitigation scheme by placing
an upper cap on the utilization of resources at routers.
Exploiting peer-to-peer application: Le Blond et al. 
suggested that peer-to-peer applications can be exploited
to trace IP addresses of users running Tor. Moreover, scan
of malicious Tor exit relays should be used to correlate
various user streams for deanonymization. Experiments
showed that their ‘bad apple’ attack was able to identify
193% more streams, including 27% HTTP streams, and
reveal IP addresses of 10,000 Tor users. This constituted
9% of all the ﬂows passing through the exit relays under
Induced Throttling Attacks: Geddes et al.  pro-
posed a new attack which breaches the Tor network by
exploiting its selection bias in favor of high capacity re-
lay nodes. Authors showed that induced throttling at the
corrupt exit node by exploiting congestion or traﬃc shap-
ing algorithms can induce similar traﬃc patterns at other
relays associated with the corrupted exit relay.
Using Decoy Traﬃc: Chakravarty et al.  used de-
coy traﬃc on anonymous networks to detect traﬃc inter-
ception. The proposed strategy is based on the idea of
injecting traﬃc containing bait credentials for decoy ser-
vices requiring user authentication. Chakravarty et al. set
up decoy IMAP and SMTP servers and identiﬁed ten in-
stances of traﬃc interception over ten months.
3.1.4. Tor Traﬃc Analysis Attacks
A few studies have focused on the analysis of Tor traf-
ﬁc for breaching this network. Analysis shows that traﬃc
analysis can provide an eﬃcient mechanism for deanonymiza-
tion. A few of these studies are summarized as follows.
Traﬃc Correlation Attacks: Johnson et al.  con-
ducted a thorough analysis of the Tor network with a deep
focus on the development of a threat model. They built
the Tor path simulator (TorPS) to assess Tor’s vulnera-
bility to correlation based attacks. Their study suggested
that a single Tor relay adversary can deanonymize 80% of
users within six months. This research showed that set
of relays is dependent upon the user’s application which
reduces security of the Tor network.
Using Timing Signature: Murdoch and Danezis 
presented a simple mechanism to evaluate the Tor nodes
being used in a circuit. In the proposed scheme, a ma-
licious node sends probe data to the Tor relays. All Tor
relays used in the circuit will experience a delay and client-
server communication will be modulated. Hence, correla-
tion between delay and modulation gives insight about the
relays being used in a circuit.
Traﬃc Analysis Attack: Chakravarty et al.  used
NetFlow data to analyze the eﬀectiveness of traﬃc anal-
ysis attacks against Tor network. Their proposed attack
creates variations in traﬃc at the server end and observes
the eﬀects at a colluding server at the other end. They
reported 81.4% accuracy in real-world experiments with
6.4% FP rates.
Queue Scheduling and Resource Allocation: Zhang et
al.  proposed a priority queue scheduling mechanism to
reduce the correlation between high load and high latency
which would ultimately increase the level of anonymity.
However, increase in anonymity comes at the cost of la-
tency which degrades quality of service at the user end.
Extensive experiments using the proposed mechanism showed
an increase in anonymity due to decrease in correlation be-
tween load and latency.
Correlation Using K-means Algorithm: Song et al. 
applied machine learning techniques to deanonymize Tor
ﬂows at the ﬁrst hop and last hop in the network. They
used the time / stream size tuple of attributes together
with the k-means algorithm to deanonymize by matching
ﬁrst hop traﬃc with last hop traﬃc. Their results showed
that as little as 8 packets are enough to deanonymize a
Tor stream with greater than 99% accuracy.
Website Fingerprinting Using Machine Learning: In
, Panchenko et al. suggested the use of machine learn-
ing approaches requiring feature selection and classiﬁca-
tion for website ﬁngerprinting. Authors used Support Vec-
tor Machine (SVM) classiﬁer with various features includ-
ing packet sizes (except 52 size packets because of excess
use in acknowledgements), packet size markers to express
direction of ﬂow, HTML markers, total transmitted bytes,
number markers, occurring packet sizes, percentage incom-
ing packets and number of packets. Extensive research
showed that volume, time and direction of the traﬃc were
the most promising features and classiﬁcation of close-
world and open-world dataset gave 55% detection rate.
However, camouﬂaging the traﬃc decreased the detection
rate to 3%.
Website Fingerprinting Using New Metrics: Wang and
Goldberg  proposed the use of Tor cells as a unit of data
transfer rather TCP/IP packets for website ﬁngerprint-
ing. Authors collected data using realistic assumptions on
adversaries from client to entry guard node. The study
suggested the removal of SENDME cells as they do not
play any signiﬁcant role in improving performance. Pro-
posed metrics use the observation that dynamic content is
present in only incoming packets and it is present at the
end of the packets. Upto 95% recall rate and 0.2% FP rate
is observed using SVM classiﬁer.
Wavelet Decomposition Attack: Jin and Wang  sug-
gested a wavelet based decomposition mechanism to esti-
mate the distortion in timing at the receiver end of Tor net-
work. Authors showed that wavelet based multi-resolution
analysis (MRA) captures the variability of the timing dis-
tortion at all levels, with better granularity than tradi-
tional estimation of timing distortion. Deanonymization
rate of 96% was obtained for Tor at a packet rate of 4
pkts/sec in 3 minutes without changing established paths
(circuits) of Tor. Analysis showed that Tor circuit rota-
tion could decrease the accuracy of deanonymization to
72% after 5 Tor circuit rotations in 3 minutes.
Exploiting Side-channels to Identify Clients: Gilad and
Herzberg  exploited three kinds of side-channels includ-
ing (1) globally incrementing IP identiﬁers, (2) packet pro-
cessing delays, and (3) bogus-congestion events. Sequen-
tial port allocation is also used to identify the clients. Two
scenarios for breaching have been presented including (1)
fully oﬀ-path attack to detect TCP connections, and (2)
detecting Tor connections by eavesdropping on clients.
3.1.5. Tor Improvements
In this section, we present some miscellaneous studies
about improvement in Tor network by focusing on Tor re-
lays, path selection mechanisms, transport layer protocols
and application layer improvements.
Misusing Tor Exit Node: Gros et al.  studied the
abuse and misuse of Tor exit nodes to compromise the
anonymity of the Tor network. Authors proposed a mech-
anism, called Honeywall, to avoid misuse of any Tor exit
node. According to Honeywall, whenever any exit node
detects a malicious behavior, it lowers the reputation of
the immediate predecessor router and also sends an alert
to it. Similarly, the intermediate router lowers the repu-
tation of its predecessor router. Through this strategy, all
“bad” nodes eventually end up with lower reputations and
all “good” nodes have higher reputation.
Exposing Exit Relays: Winter and Lindskog  de-
tected malicious exit Tor relays and proﬁled their behavior.
An exit relay scanner was built to identify all outgoing Tor
traﬃc and identify the malicious nodes and avoid man-in-
the-middle attacks. Patches were built for the Tor browser
bundle to collect certiﬁcates through multiple paths to
check authenticity of the destination server.
Tuning mechanism for Tor: Xin and Neng  showed
that Tor lacks the evaluation system for the node store.
Authors presented and theoretically analyzed a tuning mech-
anism for Tor. The proposed tuning system included the
establishment of an evaluation system and optimization of
Tor node store and output mode. Through the evaluation
system, all nodes are ranked based on their anonymity,
uptime, bandwidth and latency. In the optimization stage
authors suggested to use a ﬁxed number of circuits such
that traﬃc load has least eﬀect on latency.
Increasing Security of Tor Network: Backes et al. 
conducted research on the security of Tor network for anony-
mous browsing and presented a novel security protocol.
Authors elaborated the concept of security in anonymity
softwares. Their study showed that current key exchange
algorithms are ineﬃcient and a number of security en-
hancements were suggested including cryptographic require-
ments for secure browsing.
Transport Layer Improvements: Marks et al.  stud-
ied TCP based deﬁciencies in the Tor network. By study-
ing the transmission mechanism of Tor, authors proposed
to split bidirectional links into two separate TCP links.
Experiments with separate TCP links showed 100% in-
crease in throughput with a decrease in throughput vari-
ance from 43000 KB/s to 10000 KB/s .
Switching to uTCP and uTLS: Nowlan et al.  probed
into the cross-stream head of line blocking problem of TCP
in the Tor network. Their study suggested the use of un-
ordered TCP (uTCP) and unordered TCL (uTLS) for re-
ducing inter-dependence in inter-leaving streams, due to
the requirement of low latency in the Tor network.
Feasibility of DOS attack over Tor: Danner et al. 
conducted a deep investigation on the feasibility of Denial-
of-Service (DoS) attack (proposed in a previous study by
Borisov et al. ) over Tor network. Authors showed
through simulations and analytical evaluations that cor-
rupted relay nodes can be used to exploit Tor network
and perform DoS attack. Authors suggested the use of
reliable guard nodes (entry and exit) which can decrease
the probability of selection of a corrupt Tor relay.
3.1.6. Anonymity without Tor
To present a glimpse of studies providing anonymity
without Tor, we present a few studies focusing on packet
encapsulation and central server based anonymity mecha-
Anonymity Using a Central Server: Herzberg et al.
 proposed a camouﬂaged browsing design using a cam-
ouﬂaged server. The basic idea is to communicate with
the camouﬂaged server using a manner similar to popular
web services. Encrypted communication, URLs of GMAIL
with packet frequency and sizing similar to GMAIL can
easily pass unnoticed through any adversary. Although
this design provides better anonymity, it suﬀers from a
single point of failure.
In-Network IP Anonymization Service: Mendonca et
al.  presented a novel idea of user anonymity by work-
ing with a network service provider. Proposed service
AnonyFlow used an in-network IP anonymization service.
The fundamental idea was to conceal the source identi-
ﬁer from the other side of the network. An OpenFlow
based implementation was used for performance evalua-
tion. However, anonymity could be breached by compro-
mising the network service provider.
3.2. Tor Path Selection
Tor selects three relays based upon its path selection
algorithm which incorporates anonymity and reliability
characteristics of relays and users . By compromis-
ing the path selection mechanism, the complete anonymity
mechanism of Tor can be breached. In this subsection,
we present studies covering (1) new algorithms for path
selection, and (2) analysis of path selection algorithms.
An overview showing the classiﬁcation of ma jor studies is
shown in Figure 6.
Table 5 presents a comparison of various research works
in Tor’s path selection track. Comparison shows that per-
formance and anonymity were the most frequently studied
parameters for path selection. However, majority stud-
ies neglected autonomous systems, relay locations, hop
counts, multi-path mechanisms, load and relay capacity.
Moreover, majority research works focus on the develop-
ment of new algorithms while few studies analysed the
current path selection algorithms.
3.2.1. New Path Selection Algorithms
LASTor - Low Latency with Better Anonymity Algo-
rithm: Akhoondi et al.  proposed a new path selection
algorithm namely LASTor. LASTor incorporates the loca-
tions of relays before choosing paths and does not always
select the shortest path as it reduces the entropy of path se-
lection. Moreover, LASTor avoids paths passing through
ASs which can compromise anonymity of the system by
Figure 6: Focus of published research on Tor’s Path Selection ap-
Table 5: Research works on Tor’s path selection. Table entries symbolize New algorithms (New Algo), Analysis (Anal.), Autonomous Systems
(AS), Relay Locations (Relay loc.), Hops, Performance-Latency-Bandwidth (Perf., Lat, BW), Multi-path, Load, Relay Capacity (Rel. Cap.)
and Anonymity (Anon).
Focus Path Selection Parameters Idea
Research New Anal.AS RelayHopsPerf.Multi-LoadRel. Anon.
Algo loc. Lat,
New Path Selection Algorithms
Akhoondi et al.  X X X Included relay locations and autonomous system
Chen et al.  X X X X X Included hops and geographic distance in path selection
Karaoglu et al.  X X X Studied multipath design
Panchenko et al.  X X X X Studied Load and Capacity at nodes
Li et al.  X X X Proposed tunable mechanism varying between
anonymity and performance
Panchenko et al.  X X X X X Studied latency, link capacity and load at nodes
Liu and Wang  X X X X Proposed random walk based circuit building protocol
Liu and Wang  X X X Proposed new relay selection mechanism with backup
Snader and Borisov
X X X X Studied malicious nodes, proposed balance between
anonymity and performance
Li et al.  X X X Proposed relay recommendation system
Tang and Goldberg
X X Suggested the use of bursty circuits instead of busy
Wang et al.  X X Included latency as a measure of congestion in path
Snader and Borisov
X X X Suggested opportunistic bandwidth measurement with
priority based traﬃc handling
Elahi et al.  X X X Discouraged short term entry guard churn and
time-based entry guard rotation
Analysis of Path Selection
Bauer et al.  X X X Suggested random or Snader-Borisov approach for
Wacek et al.  X X X Suggested bandwidth weighted relay selection and
avoidance of congested circuits
Using MultiPath Routing Karaoglu et al.  evalu-
ated the multipath design for Tor network to avoid conges-
tion and overcome the limitations in Tor’s circuit construc-
tion. Evaluations revealed a four-fold increase in through-
put with better load balancing and traﬃc mixing. How-
ever, high buﬀer costs at the Tor proxies were the major
limitations of multipath design.
Path Selection Using Load and Capacity of Nodes:
Panchenko et al.  studied the delays in the Tor network
and provided new measures in path selection to improve
user experience. Two factors used for path selection design
are “load” at the nodes and maximum “capacity” at the
nodes. Authors showed that these factors can increase the
performance by 70%. Their study concludes that nodes,
not edges, are the deciding factors for performance.
Tunable Mechanism of Tor: Li et al.  emphasized
the development of a tunable mechanism for Tor users de-
pending on anonymity and performance required by users.
Authors used “path length” as a metric to tune user re-
quirements based upon anonymity and performance fol-
lowed by client side modiﬁcations of Tor protocol. Results
showed that browsing time deteriorates quickly from 14.4
to 140.1secs with a 37.3% increase in failure rate by in-
creasing the path length from 2 to 6. The proposed mech-
anism requires only client side modiﬁcation.
Using Latency and Link Capacity: Panchenko et al.
 evaluated the impact of diﬀerent factors on the per-
formance of the Tor network. Factors considered included
overloaded nodes and links and geographical diversity of
nodes. Authors presented a novel path selection algorithm
based on latency experienced by the nodes and link ca-
pacity. Metrics used for evaluations included circuit setup
duration, round trip time (RTT), stream throughput and
inﬂuence of penetration.
Random Walk Based Circuit Building Protocol: Liu
and Wang  presented a random walk based circuit
building protocol (RWCBP) which is a two-step method:
circuit construction, followed by application message trans-
mission. Network latency, computational latency and trans-
mission loads were used to analyze the performance of
the proposed protocol. Using indexes of performance and
anonymity, resilience of the proposed protocol was ana-
New Circuit Building Protocol: Liu and Wang 
studied the current protocol of the Tor network and pro-
posed a novel circuit building design with two phases: se-
lection of user selectable relay nodes and circuit construc-
tion. Authors presented enhancements in the selection of
relay nodes, fast circuit construction and backup circuit
algorithm. Better performance and user experience are
obtained with the new protocol while achieving the same
level of anonymity.
Tunable Path Selection for Better Security and Per-
formance: Snader and Borisov  addressed the issue
of the selection of malicious nodes in the path selection
due to self-advertised bandwidth. Authors proposed an
algorithm which is based upon the anonymity and perfor-
mance in the Tor network. Signiﬁcant performance gains
were observed using the proposed strategy with single and
multipath route selection.
Relay Recommendation System: Li et al.  pro-
posed a relay recommendation system to provide reliable
information about all relays for building circuits (paths).
Its main goals include the mitigation of low-resource at-
tacks, better performance and tradeoﬀs between anonymity
and performance. Authors proposed path selection algo-
rithms for increased anonymity. Signiﬁcant performance
gains with increase in anonymity were observed in the sim-
ulations of the proposed scheme.
Preferring Bursty Circuits over Busy Circuits: Tang
and Goldberg  proposed a new algorithm which sug-
gests the use of bursty circuits instead of busy circuits. Au-
thors suggest that bursty circuits (such as web browsing)
can provide less latency than the busy circuits (used for
bulk data transfer). Proposed circuit selection algorithm
uses exponentially weighted moving average (EWMA) of
cells sent on any path and uses the path with lowest EWMA
(because new and bursty paths have high EWMA).
Incorporating Congestion in Path Selection: Wang et
al.  proposed a novel path selection algorithm which
incorporates the latency of nodes as a measure for conges-
tion. The proposed algorithm favors nodes which provide
lower latency. Study suggests that node latency is greater
than the link latency in majority of the cases. Authors
conclude that the proposed algorithm can reduce latency
by upto 40%.
Opportunistic Bandwidth Measurement Algorithm: Snader
and Borisov  addressed Tor’s shortcoming of favor-
ing high bandwidth nodes based on advertised bandwidth.
Their study showed that an opportunistic measurement
of bandwidth for all routers by other connected routers
can reduce the vulnerability risk by any adversary in Tor.
Moreover, priority based traﬃc handling, i.e., high perfor-
mance or high anonymity can reduce the risk of partition-
Analyzing and Improving Entry Guard Selection: Elahi
et al.  conducted an in depth investigation on the se-
lection of entry guards in Tor network. The study showed
that short-term entry guard churn and explicit time-based
entry guard rotation result in an increased usage of en-
try guards in clients, which results in a greater number of
Trust-Aware Path Selection Algorithm: Johnson et al.
 proposed a path selection algorithm which uses the
probability based distribution to keep itself aware of the
location of adversaries in the Tor network. In develop-
ing trust based model, authors take the relays uptime as
the most trustworthy factor in determining the selection
of the path. Bypassing the paths containing adversaries
can mitigate the traﬃc analysis attacks conducted by the
Investigating Tor’s Exit Policies: Liu and Wang 
studied the exit policies of the exit nodes and addressed
the short-comings in the current Tor architecture. A new
protocol was proposed which comprised of three parts: (1)
reporting misbehavior protocol, (2) building global black-
list protocol, and (3) blocking misbehavior protocol for
users. User experience, performance and anonymity were
the key indexes used for evaluation.
3.2.2. Analysis of Path Selection
Predicting Path Compromise: Bauer et al.  showed
that the current mechanism of Tor is vulnerable to path
compromise because Tor selects paths based on bandwidth
capabilities of routers. Study shows that the application
level protocol is a signiﬁcant factor to predict path com-
promise. Research suggests that router selection should be
random or through Snader-Borisov approach to avoid any
bias in router selection. Study showed that most robust
applications for path compromise are HTTP and HTTPs
applications while the most vulnerable are peer-to-peer ap-
Optimizing Hops, Performance ﬂags and Geographic
Distance: Chen and Pasquale et al.  studied the path
selection mechanism by varying the number of hops, per-
formance ratings and changing the geographic distance be-
tween routers. Trade-oﬀs between anonymity and other
parameters (latency etc.) were extensively evaluated. The
authors concluded that reduction in hops and geographic
distance can increase throughput and decrease anonymity.
Empirical Evaluation of Relay Selection: Wacek et al.
 evaluated the relay selection mechanism of Tor to
estimate latency. Performance and anonymity were an-
alyzed for a number of relay selection techniques under
varying load conditions. The authors suggest that a com-
bination of bandwidth-weighted relay selection and avoid-
ance of congested circuits can provide better throughput
and less latency.
3.3. Tor Analysis and Performance Improvements
In this section, we cover studies on Tor dealing with its
analysis and performance improvement mechanisms. Clas-
siﬁcation of various studies is shown in Figure 7.
Table 6 presents a comparison of various generalized
studies covering the modelling of Tor network. Compar-
ison shows that majority studies focused over analysis of
Tor network. Moreover, usability analysis and anonymity
analysis were the most frequently studied topics followed
by performance analysis. Very few research works focused
over the sociability issues of Tor network.
3.3.1. General Studies of Tor
Several studies covering pros and cons of Tor and ana-
lyzing statistics of Tor referring to users’ quality of expe-
rience are summarized in the paragraphs below.
Understanding Challenges and Social Factors: In ,
Dingledine et al. described the challenges in implementa-
tion of Tor and discussed social issues. Tor network design
and its details were also discussed with reference to the
previous state-of-the-art. Possible avenues for improve-
ments in the Tor network and ﬂaws in the current system
were presented including abuse, security implications and
perceived social value.
Who is More User Friendly ? Abou-Tair et al. 
focused on the usability of diﬀerent anonymizing solutions
including Tor, I2P6, JAP/JonDo (Java Anonymous Proxy)7
and Mixmaster8. The installation of all softwares was
analyzed with regard to ease-of-use. They measured the
bandwidth consumption of all softwares. The authors con-
cluded that I2P and Mixmaster provide better anonymity
but are more complex. On the contrary, Tor and JAP
are easy to use but comprise somewhat on the degree of
anonymity they provide.
Usability Analysis of Tor: Clark et al.  conducted
usability analysis for deployment of Tor and software tools
associated with Tor including Vidalia, Privoxy, Torbutton
and FoxyProxy. Research showed that all implementations
have associated pros and cons. The study presented guide-
lines for future implementations for maximum usability of
6I2P is an anonymous overlay network which supports both TCP
and UDP traﬃc. Web: https://geti2p.net/en/
7Java Anon Proxy allows web browsing with pseudonymity using
its proxy based system. Web: https://anonymous- proxy-servers.
8Mixmaster is a Chaumian mix network which is an anony-
mous remailer providing security against traﬃc analysis and sender
deanonymization. Web: http://mixmaster.sourceforge.net/
Figure 7: Focus of various research works on the analysis of Tor.
anonymity tools. Research spanned over the installation,
conﬁguration, usage menu, veriﬁcation and switch-oﬀ fea-
tures of various anonymity tools.
Safeplug vs. Tor: Edmundson et al.  analyzed
the security provided by Safeplug in comparison to the
Tor network. Safeplug9is a plug-and-play network de-
vice which is plugged into the router and it acts as an
HTTP proxy by directing all web traﬃc through the Tor
network. Safeplug was launched to provide ease in access
for Tor users. On the contrary, Tor network can be ac-
cessed through Tor browser bundle provided by Tor. Study
showed that Safeplug was vulnerable to ﬁrst and third-
party trackers, through which users can be deanonymized.
Attacker can modify the settings of Safeplug externally
through cross-site request forgery (CSRF). Safeplug pro-
vided more latency and less protection than Tor.
Robustness of Tor: Barthe et al.  argued that ro-
bustness has always been neglected while privacy is the
issue that receives most attention. Authors deﬁned gen-
eral and ﬂexible deﬁnitions for robustness and studied the
Golle and Juels protocol. By identifying the weaknesses
in the current protocol, novel enhancements were also pro-
posed for robustness.
Anonymity and Monitoring on Tor: Mulazzani et al.
 addressed the Monitoring and Anonymity issues in
the current Tor network. A dataset was collected over a
period of six months. Analysis showed that a sinusoidal
pattern in users is observed with half of servers located in
Germany and United States. A proposed implementation
has been added into TorStatus10, which is the project dis-
playing Tor network status, available routers, bandwidths,
hosts and availability history.
Tor Traﬃc Statistics: Huber et al.  analyzed the
HTTP usage of the Tor network. Research showed that
78% of Tor users do not use Tor using TorButton, which
can be used for deanonymization. 1% of Tor requests are
vulnerable to piggybacking attacks. 7% requests, pertain-
ing to social networks, contain identiﬁable information.
The authors suggested the use of HTTPS instead of HTTP
for secure communication.
Tor Usage Statistics: McCoy et al.  studied the
applications, user countries and usage of Tor network. Statis-
tics collected from Tor showed that non-interactive proto-
cols (BitTorrent traﬃc), comprising of a minority of con-
nections, consumed majority of resources. Non-secure pro-
tocols like HTTP can be exploited by the exit router to
log sensitive information. The study suggested a protocol
for identiﬁcation of all exit routers capturing POP3 traﬃc.
Usage statistics revealed that USA, Germany and China
are major users of Tor.
Statistical Data of Tor: Loesing et al.  collected
the statistics from the live Tor network to measure two
aspects of communication, i.e., (1) country wise usage, and
(2) traﬃc port numbers for exiting traﬃc. Both these
Table 6: Research works on general study and modelling of Tor. Table entries symbolize Discussion (Dis.), Analysis (Anal.), Propose (Propos.),
Sociability Issues (Soci. Issu.), Usability Issues (Usab.), Performance Latency (Perf. Lat.), Performance - bandwidth (Perf. BW), Anonyimty
Study focus Research Parameters Idea
Research Dis. Anal.Propos.Soci. Usab. Perf. Perf. Anon.
issu. [St./eas] lat BW
General Study over Tor
Dingledine et al.
X X Discussed challenges and social issues, and studied Tor
Abou-Tair et al.
X X X X X Studied usability, bandwidth and anonymity over
Clark et al.  X X X Performed usability analysis of Tor with other anonymity
Edmundson et al.
X X X Compared anonymity and performance of Safeplug with Tor
Barthe et al.  X X X Studied robustness in Tor network
Mulazzani et al.
X X X Analysed monitoring and anonymity issues in Tor
Huber et al.  X X X Studied anonymity using HTTP usage statistics
McCoy et al.  X X X Studied applications, usage statistic and misusage of Tor
Loesing et al.  X X X Studied country and port usage of Tor
Chen et al.  X X X Proposed anonymous payments over anonymous network
Modelling Tor network
Jansen et al.  X X X X Proposed graph based Tor topology
Jansen and Hopper
X X X Developed discrete event Tor simulatork
Bauer et al.  X X X Developed emulation toolkit ExperimenTor for Tor
statistics can be used for future improvements in the Tor
network for better anonymity services. The study also
revealed that port 80 receives most traﬃc.
Micropayments Using Tor: Chen et al.  proposed
a novel mechanism of anonymous payments for network
services. The proposed mechanism allows users to make
untraceable micro-payments to each other. Authors in-
cluded features of oﬄine veriﬁcation, overspending preven-
tion, aggregation and low overheads. Experiments showed
only 4% overhead for the proposed strategy.
3.3.2. Modeling Tor Network
In this section, we present the modeling techniques
used for analyzing Tor.
Modeling Topology and Hosts of Tor: Jansen et al.
 developed a model of Tor which closely resembled
the Tor network. Authors developed a graph for Tor topol-
ogy where vertexes related to downstream bandwidth, up-
stream bandwidth and packet loss, and edges related to la-
tency, jitter and packet loss. All hosts including relays, au-
thorities, clients and Internet servers were mapped to the
developed graph based on characteristics obtained from
Shadow: Simulating Tor Network Jansen and Hopper
 developed an open source discrete event simulator for
simulating the network layer of Tor on a single machine.
Authors compared the performance of Shadow with real-
world simulation results from the PlanetLab testbed.
Emulation Toolkit for Tor Experimentation: Bauer et
al.  developed ExperimenTor, an emulation toolkit
for Tor network. Their research was focused on the toolkit
rather than the analysis of the Tor network.
3.3.3. Analysis of Tor
Analysis of Tor network has been a part of many stud-
ies covering delays, bandwidth, quality of service, relay se-
lection and authentication protocols. Several studies cov-
ering these areas are summarized in following sections.
Table 7 presents a comparison of various research works
in analysis and performance improvement track. Compar-
ison shows that relay selection and latency analysis are
the most frequently studied topics followed by anonymity,
bandwidth and quality of service analysis. Very few stud-
ies focused over queues, traﬃc shaping techniques and pro-
Understanding Delays in Tor: Dhungel et al. 
analyzed the delays in the entire Tor network. Authors
suggested that overlay network plays the most signiﬁcant
role in Tor. The study revealed that 11% of Tor routers are
overloaded with traﬃc which resulted in very high delays.
In 7.5% of circuits, overall latency introduced a 450ms
delays. Guard routers incorporate more delay than non-
guard routers. There is high ﬂuctuation in delay for all
routers except for those having high bandwidths.
Measurement and Statistics: Loesing et al.  stud-
ied the latencies inside the Tor network. A deep investiga-
tion was conducted to evaluate the individual delays and
QoS properties. The authors showed that circuit building
time (Introduction and Rendezvous) is the most crucial
Table 7: Research studies on analysis and performance improvements of Tor. Table entries symbolize New algorithms (New Algo), Analysis
(Anal.), Relay Selection (Relay Sel.), Performance Latency (Perf. Lat.), Performance Bandwidth (Perf. BW.), Quality of Service (QoS),
Queues, Protocol Messages (Prot. Msgs.), Traﬃc Shaping (Traﬀ. Shap.), Anonymity (Anon.).
Focus Path Selection Parameters
IdeaNew Anal.Relay Perf.Perf. QoS Queues Prot.Traﬀ. Anon.
Algo Sel. Lat. BW. Msgs.Shap.
Analysis of Tor network
Dhungel et al.  X X X Analysed latency for Tor relays
Loesing et al.  X X X Analysed latency, QoS and performance of Tor
Ehlert et al.  X X X Studied bandwidth and latency in Tor
Pries et al.  X X X Investigated bandwidth for various path selection
Liu and Wang  X X X X X Proposed relay reliability mechanism considering
performance anonymity and QoS
Wang et al.  X X Performed an empirical analysis over family nodes
X X X Proposed fairness model for eﬃcient and fair
Chaabane et al.
X X Studied misuse of Tor’s exit nodes as proxies
Hopper et al.  X X Analysed Tor’s performance considering key
Lenhard et al.  X X Studied communication overhead in low bandwidth
networks for Tor’s hidden services
Goldberg  X X X Analysed anonymity with Tor;s authentication
Tor performance improvement
Jansen et al.  X X Proposed token based performance mechanisms for
recruiting more relays
Dingledine et al.
X X Proposed priority based traﬃc handling for relays
Wang et al.  X X X Proposed node reliability mechanism to avoid
blockage of bridges
Smits et al.  X X X X Proposed packet authorization based mechanism to
protect bridges from eavesdroppers
Moghaddam et al.
X X X Proposed traﬃc morphing (using Skype traﬃc) to
Weinberg et al.
X X X Proposed traﬃc shaping (by assembling regular
HTTP traﬃc) to avoid deanonymization
X X Suggested latency reduction by separate TCP
connections for interactive and bulk traﬃc
AlSabah et al.
X X X Proposed traﬃc morphing (using Skype traﬃc) to
Jansen et al.  X X X Proposed throttling mechanisms for reducing
latency by avoiding bulk traﬃc
delay period in Tor. Fr´echet and exponential distributions
were combined to analyze the response times.
Comparing Bandwidth with Latency: Ehlert  com-
pared the bandwidth and latency performance of Tor net-
work with the popular I2P network. Authors measured
the core latency (HTTP GET requests durations), aver-
age latency (webpage download times including external
threads and pictures) and bandwidth (download speeds).
This research showed that I2P network provides lower core
latency and Tor network excels in average latency and
bandwidth, owing to the nodes distribution and penetra-
tion of the Tor network.
Tor QoS with Path Selection Strategy: Pries et al. 
suggested that TCP suﬀers severe performance degrada-
tion from the random path selection of Tor. Slight QoS
improvement is achieved with Tor’s bandwidth weighted
path selection algorithm. The main reason attributed for
small improvements is low bandwidth of Tor routers.
Behavior of Family Nodes: Wang et al.  pre-
sented an empirical analysis of Tor family nodes. A rich
dataset of live Tor network comprising of three years was
used to study the impact of family nodes. The study sug-
gested that family nodes provide stable and better service
than other nodes. Moreover, attacks on family nodes can
disrupt the Tor network more severely than random Tor
Fairness in Tor: Tschorsch and Scheuermann  an-
alyzed the fairness issues in the current Tor network. Large
unfairness was observed in the current resource allocation
mechanism of the Tor network. Authors proposed a max-
min fairness based model for eﬃcient and fair scheduling
of resources. The proposed design was analyzed with Tor’s
N23 congestion feedback mechanism.
Misuse of Tor: Chaabane et al.  showed that Tor
network was being used for transmitting P2P traﬃc (Bit
torrent etc.) over the Tor network. HTTP and Bit torrent
were analyzed on the Tor network. The study showed that
Tor exit nodes are being used as one hop SOCKS proxies
through tunneling. New techniques were devised to de-
tect such abnormalities in exit nodes’ behavior. Research
showed that simple crawling over exit nodes can be used
to collect as many bridge identities as needed.
Challenges for Hidden Services of Tor: Hopper 
conducted research on the botnet attacks on Tor through
hidden services. Tor exhibits poor network performance
due to increased load on relays under such attacks. Hop-
per attributed the poor performance to the key exchange
mechanism of Tor. Study showed the possible research di-
mensions of limiting request rates from botnets, throttling
entry guard, reusing failed partial circuits and isolating
hidden services circuits.
Tor Hidden Services in Low Bandwidth Access Net-
works: Lenhard et al.  conducted a measurement and
statistical analysis for estimating the communication over-
head of Tor hidden services in low bandwidth access net-
works. Research showed that boot strapping time, RTT
and circuit building time were the major bottlenecks to
performance. Due to numerous delays, an increase in time-
out value was suggested to avoid repeated retransmissions.
Analysis of Tor Authentication Protocol: Goldberg 
analyzed the security of Tor’s authentication protocol (TAP).
The authors argued that any security breach by a sin-
gle malicious Tor relay can deanonymize users’ sessions.
Through empirical evaluations, research showed that TAP
is secure in random oracle model.
Statistics Collection Mechanism of Tor: Mani and Sherr
 analysed the data collection mechanism of Tor through
‘PrivEx’. They showed that statistics of PrivEx can be
easily compromised by the present of adversary nodes in
the Tor network. As a result of shortcomings of PrivEx,
authors proposed ‘HisTor’, a privacy preserving statistics
collection mechanism of Tor which is much more diverse
than PrivEx. HisTor uses the count of queries by exit
nodes and relays in form of a histogram where individual
nodes have little control over the aggregate statistics.
3.3.4. Tor Performance Improvement
Owing to the increasing demand for Tor, various stud-
ies have proposed performance improvements to cope with
future demands. In this section, we present these stud-
ies covering Tor node selection, traﬃc distribution and la-
tency management etc.
Node Recruitment for Tor: Jansen et al.  focused
their research on recruitment of new Tor relays, motivated
by the fact that only 1.5% nodes participate as relays.
Authors proposed BRAIDS which is a token based mecha-
nism providing high bandwidth to those users who employ
BRAIDS. Proposed scheme characterizes traﬃc into high
throughput, low latency and normal traﬃc. Based upon
usage of BRAIDS and node networking stats, tickets are
generated which can be used to increase bandwidth.
Encouraging Tor nodes for traﬃc relaying: Dingledine
et al.  proposed a mechanism to encourage Tor nodes
for traﬃc relaying. Study suggested a priority based traﬃc
handling, which gives more weight (in form of bandwidth
and delays) to those nodes contributing resources to Tor.
However, all Tor relays carry an additional load of priority
based traﬃc handling. Directory authorities need to assign
priority levels to all Tor users participating in Tor relays.
Improving Distribution Mechanism of Tor Bridges: Wang
et al.  improved the distribution mechanism of Tor
bridges by implementing node reliability statistics to avoid
the blockage of bridges by corrupt nodes. The uptime of
assigned bridges is used to give reputation points to users.
In case of any blockage of a bridge, a new bridge address is
given on payment of earned credit. To ensure anonymity,
reputation information is stored on users’ systems by using
a privacy-preserving technique which cannot be circum-
vented by malicious users.
Packet Authorization for Tor Bridges: Smits et al.
 proposed BridgeSPA, a packet authorization based
mechanism, to protect users of Tor hosting Bridges. All
Tor user hosting bridges are susceptible to traﬃc analysis
attacks. To counter this attack, the authors suggest the
transmission of a bridge key by bridge distribution author-
ities which is valid for a limited time, as determined by the
bridge. For any communication with the bridge, Tor users
should use that key within the assigned time period.
SkypeMorph - Tor traﬃc Shaping: Moghaddam et al.
 proposed a new mechanism namely SkypeMorph to
avoid the censorship of Tor bridges. The fundamental idea
was to hide Tor traﬃc as Skype video traﬃc (a widely used
protocol). SkypeMorph, which runs side by side with Tor,
makes it hard to distinguish Tor traﬃc from Skype traf-
ﬁc. Two schemes were suggested for traﬃc morphing, (1)
using the target stream attributes, (2) incorporating both
source and destination streams to incorporate packet tim-
ings. Both streams provided nearly identical performance,
but the former had lower complexity.
StegoTorus - Steganographing Tor Traﬃc: Weinberg et
al.  proposed a novel technique to bypass censorship
on Tor. Their scheme is based upon the idea of chopping
Tor traﬃc into multiple streams, resembling HTTP traﬃc,
before passing through the censor. StegoTorus acted as a
proxy on Tor clients.
Torchestra - Separate connections for Interactive and
Bulk Traﬃc: Gopal and Heninger  proposed the trans-
mission of interactive and bulk traﬃc over two separate
TCP connections among all nodes in the Tor network. Ex-
ponentially weighted moving average (EWMA) algorithm
was used to distinguish between interactive and bulk traﬃc
on all circuits. Upto 40% reduction in delays was observed
as compared to standard Tor for the proposed strategy.
Reducing Latency in Tor: AlSabah et al.  pro-
posed a mechanism for congestion control and ﬂow control
in order to reduce latency in the Tor network. The study
suggested the use of small ﬁxed size windows and small dy-
namic windows which can reduce the packets in ﬂight. For
ﬂow control, the study proposed an N23 algorithm which
caps the queue lengths of Tor routers and provided a 65%
increase in webpage responses and 32% decrease in page
Throttling Tor bulk users: Jansen et al.  ad-
dressed the poor performance of Tor network using bulk
data transfers. Three dynamic throttling algorithms were
proposed for reducing network congestion and latency. The
guard relay capped the bandwidth capacity of nodes, so,
only local relay information was used. Simulations showed
that throttling reduces the web page latency and increases
the anonymity of Tor network.
3.3.5. Tor Client Mobility
In this section, we study the research works focused
on the mobility of Tor network with a particular emphasis
on anonymity. Table 8 shows the research works in path
selection track and shows that performance and anonymity
have been the most frequently studied parameters. Details
are presented in below paragraphs.
Using Bridge Relays: Doswell et al.  analyzed
the performance of Tor for wireless devices roaming across
multiple networks. Analysis showed that the speed of mo-
bile wireless devices signiﬁcantly aﬀects the circuit build-
ing time and Tor’s performance. Authors studied the use
of bridge relays to provide persistent Tor connections for
New Architectural Designs: Andersson et al.  pro-
posed several new architectural designs for a mobile Tor
network. A trade-oﬀ between anonymity and performance
was evaluated. Several criteria used in performance esti-
mation included usability, availability, trust and practical-
ity. The study concluded that the single Tor client option
oﬀers lowest degree of anonymity.
4. Platforms for Tor Research
In this section, we study the platforms used to study
Tor network. Our observations spanning over decades of
anonymity research shows that all research works have
studied the Tor network using three diﬀerent techniques,
(1) Experiment, (2), Simulations, and (3) Analysis. Figure
8 shows that 60% of the studies used in this paper con-
ducted experiments. Only 27% of the studies conducted
experiments. In the experiment section, majority stud-
ies developed their own testbed followed by experiments
on cloud services and PlanetLab testbeds. In the sim-
ulations section, majority research works used extensive
simulations to study Tor network. Finally, some studies
analyze Tor network by collecting statistics and discussing
the sociability and usability issues of Tor network. These
three classiﬁcation categories are elaborated in Figure 9
which shows the platforms used to study Tor network.
4.1. Tor Experiments
Studies covering Tor experiments have focused over
several areas including (1) private setup establishment, (2)
PlanetLab experiments, (3) cloud services (4) OpenFlow
networks and (5) universal composability framework.
Table 9 presents the clients, relays, servers, Tor services
and Tor implementation used by various research works.
Comparison shows that majority studies deployed their
private testbeds with 1-2 clients and 1-2 servers. Several
studies deployed limited number of relays for experiments.
Number of clients were increased drastically in the Planet-
Lab and cloud setup for Tor experiments. Moreover, traﬃc
Private Setup 86
Planet Lab 5
Cloud Services 5
UC Framework 2
New Simulator 75
Shadow Simulator 8
Figure 8: Classiﬁcation of platforms for Tor’s research.
Table 8: Research works on Tor’s client mobility. Table entries symbolize New algorithms (New Algo), Analysis (Anal.), Autonomous Systems
(AS), Relay Locations (Relay loc.), Hops, Performance-Latency-Bandwidth (Perf., Lat, BW), Multi-path, Load, Relay Capacity (Rel. Cap.)
and Anonymity (Anon).
Focus Path Selection Parameters
IdeaNew Anal.AS RelayHopsPerf. Multi-load Rel. Anon.
Algo Loc. Lat,
Doswell et al.  X X Suggested bridge relays to avoid bandwidth issues while
Andersson et al.
X X X Proposed trade-of between anonymity and performance
Figure 9: Taxonomy of platforms employed in Tor research.
analysis was the most frequently studied topic. Majority
research works used the default Tor setup without any
modiﬁcations. Figure 10 shows the classiﬁcations of ex-
periments on Tor. Analysis of ﬁgure shows that majority
of studies deployed their own private testbeds.
4.1.1. Private Setup connected with Tor
Overlier and Syverson  performed experiments by
setting up two nodes (one in Europe and other in US)
running hidden services at two ends of the Tor network.
Access to webpages and images was provided using these
services. The client PC was setup both as a client and
a middleman node, and all sampling takes place at this
Andersson and Panchenko  performed experiments
to verify the performance of their proposed mobile pro-
tocol. Mobile Tor was setup on a laptop connected to
the Tor network. The content server hosting the ﬁles
was placed at Karlstadt University. Experiments used
OnionCoﬀee, which is a Java project developed under the
Panchenko et al.  performed experiments using a
Pentium Dual Core 1GHz CPU with 2GB RAM as a client
nodes. Two existing Tor implementations (default Tor and
OnionCoﬀee) were used on the client nodes. The Internet
connection had a 10Gbps bandwidth while the local back-
bone was 100Gbps. Actual Tor relays were used to analyze
Pries et al.  performed experiments by download-
ing a 458kB ﬁle from a school web server. Command line
utility wget was used as the downloading tool. wget’s http-
proxy and ftp-proxy were conﬁgured to download all ﬁles
through Privoxy from the server. Tor release 0.1.1.26
was conﬁgured on the exit and entry nodes.
Wagner et al.  implemented a novel architecture
using Tor. Three machines were setup running Tor exit
Table 9: Experimental setups used in diﬀerent research works.
Research Servers Relays Clients Service Tor implementation
Overlier and Syverson  2 1 Hidden service default Tor
Andersson and Panchenko
1 1 Mobile Tor Onion Coﬀea
Panchenko et al.  1 Download Service Def. Tor + Onion Coﬀea
Pries et al.  1 1 Download Service Privoxy
Wagner et al.  2 1 1 Log processing WebProxy
Chan-Tin et al.  1 2 Traﬃc Analysis Def. Tor
Pries et al.  1 2 1 TCP data coll. Tor mod.
Herzberg et al.  1 Web page download Def. Tor
Bauer et al.  6+ 1+ Path compromise Def. Tor
Song et al.  6 1 Traﬃc Analysis Tor mod.
Dhungel and Steiner  1 2 1 Traﬃc Analysis Def. Tor
Gros et al.  2+ Traﬃc Analysis Honeywall
Wang et al.  2 Traﬃc Analysis Privoxy
Zhang et al.  1 3 1 Hidden Service Polipo
Loesing et al.  1 1+ Access Attempt Def. Tor
Chen and Pasquale  1 10 Download Def. Tor
Panchenko et al.  1+ 1+ Traﬃc Analysis Def. Tor
Houmansadr et al.  4 3 Traﬃc Analysis −
Li et al.  1 1 Download Analysis Def. Tor
Chakravarti et al.  1 2 Download Analysis Def. Tor
Mulazzani et al.  1+ Traﬃc Analysis Tor Status
Chaabane et al.  1+ 6 1+ Traﬃc Analysis Def. Tor
Bai et al.  2 6 Traﬃc Analysis Def. Tor
Barker et al.  3 15 1+ Traﬃc Analysis Def. Tor
Marks et al.  3 3 Download Analysis
Jin and Wang  1 1 Traﬃc Analysis Tor mod.
Tang and Goldberg  1 1 1 Download Analysis Def. Tor
Alsabah et al.  1 (3 Apps) Traﬃc Analysis Def. Tor
Moghaddam et al.  2+ Traﬃc Analysis SkypeMorph
Weinberg et al.  1 1 Download Analysis StegoTorus
Evans et al.  1 Traﬃc Analysis Def. Tor
Wang et al.  1 Traﬃc Analysis Def./Mod. Tor
Ehlert  1 Traﬃc Analysis Def. Tor
Barbera at al.  2 4 Traﬃc Analysis Def. Tor
Winter and Lindskog  2 2+ Traﬃc Analysis Tor mod.
Edmundson et al.  1 Download Analysis Def. Tor
Huber et al.  1 Traﬃc Analysis Def. Tor
Blond et al.  6 1+ Traﬃc Analysis Def. Tor
Lenhard et al.  1+ Hidden Service Def. Tor
McCoy et al.  3 Traﬃc Analysis Tor mod.
Chakravarty et al.  1 Traﬃc Analysis Def. Tor
Snader and Borisov  1 1 Traﬃc Analysis Tunable Tor + Vanilla
Gilad and Herzberg  1 1 Traﬃc Analysis Def. Tor
Loesing et al.  1+ Traﬃc Analysis Def. Tor
Chen et al.  3+ (VMs) 2+ (VMs) Traﬃc Analysis Def. Tor
Panchenko et al.  1+ Traﬃc Analysis Def. Tor
Wang and Goldberg  200 cores Traﬃc Analysis Def. Tor
Akhoondi et al.  50 Traﬃc Analysis LASTor
Murdoch and Danezis  1 2 Traﬃc Analysis Tor Mod.
Bauer et al.  6 2-6 40-90 40 node network
6 3-6 60-90 60 node network
Cloud Setup (Amazon EC2)
Sulaiman and Zhioua  1 Traﬃc Analysis Def./Mod Tor
Karaoglu et al.  1 4 Traﬃc Analysis Def. Tor
Biryukov et al.  50 Hidden Services Def. Tor
node, BIND (DNS server with tcpdump), and Apache
webserver, respectively. All machines were synchronised
by NTP. Connected to Tor network, WebProxy was im-
plemented in Perl. iptables was used to re-route traﬃc
from Tor exit node to Perl proxy server. All processing of
web server logs and proxy logs was performed using Perl,
sqlite and modiﬁed tcpick.
Chan-Tin et al.  setup a limited network for prob-
ing Tor network using client, burst server and probe ma-
chines. Entry and middle routers were chosen randomly
while exit node was forced by choice. Four Tor relays were
probed for the experiment and data of probes was col-
lected after every 5secs. Five connections were setup by
the client using multi-threading.
Pries et al.  setup client, server, entry malicious
router and exit malicious router by setting up four devices.
A TCP client application was built which sent and received
TCP data. Test server used port 41 and received and
displayed data on the screen. The client used tsocks to
transport its TCP stream through onion proxy. The Tor
conﬁguration ﬁle was conﬁgured to select designated Tor
entry and exit routers.
Herzberg et al.  implemented their proposed cam-
ouﬂaged browsing design over a test machine with an ADSL
connection to the Internet with 1,269kBps downlink and
103 kBps uplink bandwidth. Four diﬀerent URLs were
tested with 100 measurements and access time for browsers
was recorded. wget was used to download web pages.
Bauer et al.  built an extensive experimental setup
by establishing circuits for diﬀerent kinds of applications
with a number of malicious routers. The simulator gener-
ated 10,000 circuits with 6 to 106 malicious routers. The
path compromise rate for diﬀerent applications was esti-
mated by the selection of malicious routers.
Song et al.  used an Au3 script to capture Tor
traﬃc. Six nodes located at distinct places (India, Roma-
nia, Luxemburg, New Zealand, Chile, and Russia) were
deployed as exit nodes. Onion proxy running on a local
PC was conﬁgured to use the deployed exit nodes. Traﬃc
of all routers was captured to analysis.
Dhungel and Steiner  measured delay of Tor net-
Private Setup 86
Planet Lab 5
Cloud Services 5
UC Framework 2
New Simulator 75
Shadow Simulator 8
Figure 10: Classiﬁcation of platforms used in experimental Tor re-
work by setting up two relays instead of three. Client,
exit router and destination server were ﬁxed while the en-
try router was selected from the list of available routers.
To cope with varying network characteristics, the experi-
ment was repeated for eight months with each duration of
40 minutes. All 1,426 available routers were pinged ﬁve
times for measurements.
Gros et al.  performed experiments by using the
proposed Honeywall mechanism. All vulnerable clients
using Tor were placed on one side of the Honeywall and
Internet cloud was present on the other side of the Honey-
wall. All Tor clients had distinct private addresses while
Honeywall had a single public IP address.
Wang et al.  conducted experiments in a partially
controlled environment. The OP code was modiﬁed to
use the designated entry and exit Tor routers. Entry and
exit Tor nodes were conﬁgured to record the data relayed
through them. Internet Explorer was used at the Tor
client through Privoxy. The middle Tor router was se-
lected through Tor router selection algorithms.
Zhang et al.  used Mozilla Firefox on Fedora 11 to
access the hidden service using Polipo. The hidden server
was conﬁgured to use the bridge whose traﬃc was being
logged continuously. Clients and bridges were conﬁgured
to record the circuit ID, command, stream ID and arrival
Loesing et al.  conducted experiments on Tor by
conﬁguring the Tor client to use ﬁxed ﬁrst / entry relay,
which was being monitored continuously. Second and third
Tor relays were chosen randomly by Tor’s router selection
algorithm. A single access attempt was performed by cre-
ating new Tor clients after every ﬁve minutes over a 72
Chen and Pasquale  analyzed the throughput by
downloading a 100kB ﬁle through nearly 100 unique paths
with 10 times repeated downloads over each path. 10 Tor
clients were conﬁgured over PlanetLab testbed distributed
around the globe. A ﬁle server containing 100kB ﬁle was
hosted in the US using thttpd.cURL was used to conduct
downloads. Python was used to write measurement scripts
using the TorCtl library for Tor control port. Tor circuits
were conﬁgured to be replaced after every 30 minutes in-
stead of 10 minutes so that no middle replacement takes
Panchenko et al.  performed experiments in the
current Tor network with the estimation of delay and through-
put. In ﬁrst experiment, onion routers are ﬁxed but links
connecting the circuit are variable. 2,000 sets were built
with random onion routers. In the second experiment,
ICMP Ping was used to measure the delay between send-
ing a SYN and receiving a SYN-ACK packet. Each ping
was iterated 20 times to calculate the mean value.
Houmansadr et al.  conducted a deep experimen-
tal investigation on the Tor network. Application layer
softwares (Skype, CensorSpoofer) were executed in Virtu-
alBox virtual machines (VMs) on a Funtoo Linux machine.
Various VMs were connected through virtual distributed
Ethernet (VDE). Authors built their own plugin for VDE
which could drop packets at variable rates and also modify
packet contents. Various VDE switches were connected to
the central switch which provides DHCP connectivity to
Li et al.  tested their proposed tunable mechanism
of Tor (TMT) over the real Tor network. Two virtual
private servers (VPSes), acting as client and server, were
conﬁgured on Linode. The client was conﬁgured with the
TMT enhancement while the server hosted a web page.
Time to load the ﬁle, number of attempts and number
of failure attempts were measured to estimate the perfor-
mance of TMT.
Chakravarti et al.  setup their own client, server
and probing host machine at three distinct locations inside
US. 100MB ﬁle was placed at the web server which gave
suﬃcient downloading time to the client. Linux traﬃc
controller was used to shape the client-server bandwidth.
26 distinct Tor circuits were created and probed through
diﬀerent locations and compromised links were detected.
Mulazzani et al.  collect data by using TorSta-
tus and updating its script tns-update.pl and network-
history.php.RRDtool was used to store the values in a
round robin database (RRD). The collected dataset was
used for basic network monitoring.
Chaabane et al.  conduct a deep traﬃc analysis of
Tor using HTTP and Bit Torrent protocols. The authors
created and monitored six Tor relay nodes (placed in US,
Germany, France, Japan, Taiwan) advertising 100kB avail-
able bandwidth for 23 days. On average 20GB of data is
provided by each server on every day. Data was collected
at entry and exit relays.
Bai et al.  setup eight PCs with one PC running
Tor and one PC running java anonymous proxy (JAP).
Dummy traﬃc was generated from the other six PCs. Traf-
ﬁc was captured through ethereal.Winsock Packet Ed-
itor was used to record packets generated by a speciﬁc
application. Duration of the test was about 120mins with
Barker et al.  collected Tor network traces by devel-
oping a complete Tor setup. Firefox running on Ubuntu
was used on all machines. Using the Selenium browser
testing framework, 170 simulations were executed by ac-
cessing 30 websites. Three directory servers with 15 re-
lays were conﬁgured to be used for experiments. Regular
HTTPs traﬃc and HTTP and HTTPs traﬃc through pri-
vate Tor network were collected.
Marks et al.  conducted a simple experiment us-
ing three PCs running Linux kernel (2.6.26 and CUBIC
TCP). All three machines were connected via an Ether-
net switch. All Ethernet interfaces were conﬁgured to be
10Mbps full-duplex links. The ﬁrst and last two devices
setup TCP connections. First device sent data to the sec-
ond for 250secs while the second retransmitted after 50secs
delay for a duration of 250secs.
Jin and Wang  conducted extensive experiments by
monitoring both anonymous traﬃc and Tor traﬃc dur-
ing two experiments. In the ﬁrst experiment, an Apache
webserver on a Dell PC using Redhat Enterprise 4 Linux
was conﬁgured. A watermark encoder was installed on the
Apache proxy. A Dell Precision 390 was conﬁgured as a
NAT router to route traﬃc between client and the anony-
mous server. In the second experiment, an SSH server and
watermark encoder were installed on one machine acting as
server. SSH client and watermark decoder were installed
on another machine. Three random characters were sent
every second from one machine to the other through Tor.
Entry and exit relays were ﬁxed.
Tang and Goldberg  setup their own node (act-
ing as the middle node). Entry and exit nodes were se-
lected from the Tor nodes of the directory server. Au-
thors avoided the use of PlanetLab testbed because ma-
jority nodes were providing only 100KB/s. Webfetch was
used to download the target ﬁle (87KB) from author’s web
server. Connecting circuits and load was varied to verify
the proposed path selection strategy.
Alsabah et al.  performed real world experiments
by collecting oﬄine data of 200 circuits from three dis-
tinct application traces. All three applications (BitTor-
rent client, web browsing client and stream client) were
setup on the same machine which was conﬁgured to use
a speciﬁed Tor node as the entry node. All 200 circuits
included browsing (122), BitTorrent (49) and streaming
circuits (28). All applications collected 24 hours of data
over a 6 week period with periodic intervals.
Moghaddam et al.  implemented their proposed
SkypeMorph technique on Linux using C and C++ with
boost libraries. Authors collected traces of Skype data
set for modeling using multiple machines. The proposed
SkypeMorph scheme was tested by downloading multiple
ﬁles with and without it.
Weinberg et al.  implemented the proposed scheme
StegoTorus by deploying an experimental setup. The client
was a desktop PC in California with DSL link to the In-
ternet (downstream 5.6Mb/s, upstream 0.7Mb/s) and the
virtual host was situated in New Jersey inside a commer-
cial data center. 1MB ﬁles were downloaded over several
trials to test the performance.
Evans et al.  performed experiments on the real
Tor network for their proposed congestion attack. The
browser. Entry node was ﬁxed but the other two Tor relay
nodes were selected at random (by Tor’s router selection
Wang et al.  measured the network delays during
congestion by collecting delay readings of all Tor routers
for 72 hours in August 2011. At the next stage, authors
collected RTT measurements of the modiﬁed and unmod-
iﬁed Tor client to setup 255 circuits. In all experiments,
client machines were modiﬁed to incorporate the proposed
algorithm and measure the delay.
Ehlert  measured the performance of I2P and Tor
network. For I2P network, experimental setup consisted of
two machines, acting as dedicated outproxy and client. 500
most visited websites were used for downloading webpages.
For Tor, a client machine was connected to the Tor network
and performance parameters were measured similar to I2P
Barbera at al.  conducted controlled experiments
by setting up 100Mb/s network connected to four hosts
(possessing 2.66GHz Core 2 Duo CPUs). For real time net-
work experiments, the authors used their two OR nodes,
acting as Tor relays. CellFlood attacks were performed on
these routers and performance of attack and mitigation
scheme was analyzed.
Winter and Lindskog  deployed one relay in Russia
and two bridges in Singapore and Sweden. Multiple clients
were present in China for connection setup to Tor through
designated bridges and relays. In Singapore, a Tor relay
was hosted in an Amazon EC2 cloud. Bridge and relay
in Sweden and Russia were hosted by an institution and
data center, respectively. For vantage points in China, 32
SOCKS proxies and a VPS running Linux was used.
Edmundson et al.  analyzed the security of Safe-
plug and Tor by conducting separate experiments for both
applications. Authors measured the latency of the sys-
tem, and investigated the eﬀect of cookies and third party
trackers over both applications.
Huber et al.  deployed a Tor exit node which logs
the HTTP requests. Nine million HTTP requests were
recorded in several weeks. All requests were analyzed for
available patterns and statistics were presented in the re-
Blond et al.  conducted experiments by deploying
Tor exit nodes. Authors instrumented and monitored six
Tor nodes for a period of three weeks. One exit node was
conﬁgured to accept TCP connection for Bit torrent, in
order to perform the hijacking attack.
Lenhard et al.  ran Tor processes on their de-
vices connected to the Tor network. The hidden services
were accessed through low bandwidth access network edge.
A modem provided a data rate of 56kb/s downstream
and 44kb/s upstream. For EDGE, data rate was around
230kb/s. The broadband network provided 100Mb/s.
McCoy et al.  setup their router connected to
1Gb/s network link with a rank of top 5% Tor routers
and ﬂagged as Running. At most 20bytes were logged to
avoid information breaching laws. Setup was conﬁgured
for both experiments separately covering (1) exit router
and (2) non-exit router. Entrance and middle router traﬃc
was logged for 15 days comprising of time stamp, previous
hop’s IP, TCP port, next hop’s IP and circuit identiﬁer.
For exit traﬃc logging, tcpdump was used over the router
which relayed 709GB of traﬃc and only the ﬁrst 150bytes
of packet were logged. Ethereal was used for protocol
Chakravarty et al.  transmitted decoy traﬃc over
a custom client supporting IMAP and SMTP protocols.
The client was implemented using Perl and service pro-
tocol emulation was provided by Net::IMAPClient and
Net::SMTP. The client hosted on Intel Xeon CPU run-
ning Ubuntu Server Linux v8.04.
Snader and Borisov  performed experiments on
Tor by downloading 1MB ﬁles over HTTP connections
through various exit routers. All other entities includ-
ing guard routers, client and web server remain ﬁxed for
the entire duration of the experiment. 20,000 and 40,000
trials were performed for tunable Tor and standard Tor
respectively spanning a duration of two months.
Gilad and Herzberg  conducted an empirical inves-
tigation for the performance of proposed attacks in the Tor
network. Indirect rate reduction attack was evaluated by
experiments in the live network. For experiments, a Linux
machine ran an Apache web server. Data at the rate of
0.5KBps was transmitted.
Loesing et al.  collected Tor statistics by follow-
ing the legal requirements, user privacy, ethical approvals,
informed consent and community acceptance. Authors
collected data from the Tor network and evaluated the
port numbers and country of origin of the obtained IP
Chen et al.  developed ORPay which uses out-of-
band communication for payment primitives and control
messages. The “bank” was built using C language and
OpenSSL for encryption. Authors performed controlled
experiments consisting of a set of interconnected PCs run-
ning directory servers and Tor routers on VMs. Inter-client
bandwidth was 500 −600KB/s with 1 −2ms average la-
tency and 0.5ms for inter-VMs on the same machine. One
micropayment was made for every 20 packets.
Panchenko et al.  using standard PCs for fetch-
ing websites using Firefox with disabled active components
(Java, Flash etc.) and Chickenfoot used as the default plu-
gin. The closed-world dataset was collected from previous
studies, to obtain labeled ground truth dataset.
Wang and Goldberg  performed experiments on
SHARCNET, a parallel computing cluster. Upto 200 cores
were used for computation of SVM kernel matrix. torrc
was conﬁgured to close the circuits manually instead of
ﬁxed 10mins duration and ﬁxed entry guard selection was
disabled. iMacros and Tor controller was used to automate
site accesses. For closed world circuits, ﬁngerprinting was
performed on 100 sites with 40 instances each and using 10-
fold cross validation. For open-world experiments, Alexa’s
top 1,000 sites list was used.
4.1.2. PlanetLab Experiments
Akhoondi et al.  performed experiments in the real
Tor network by modifying the Tor Client with their pro-
posed LASTor protocol. LASTor is a Java application con-
trolling the Tor client through Control Port. 50 PlanetLab
nodes running LASTor were used as Tor clients to access
top 200 websites. Both latency and anonymization were
tested by collecting the traces of data set at the client
Murdoch and Danezis  performed experiments on
the Tor network by setting up a probe PC. A modiﬁed
version of Tor was used in the probe PC to choose routes of
length one. A TCP client was also established at the node
which connects to the SOCKS interface of Tor using socat.
Original Tor relays were used with a corrupt destination
Tor server recording the traﬃc traces. The probe server
ran at the University of Cambridge Computer Laboratory
while victim and corrupt server were run on PlanetLab
nodes. Data from 13 probing Tor nodes was collected and
analyzed in GNU R.
Bauer et al.  performed experiments over Planet-
Lab testbed by setting up two independent node networks
comprising of 40 and 60 nodes, respectively. Two and six
malicious nodes were added in the 40 node network while
three and six malicious nodes were added in the 60 node
network. Traﬃc was generated by six machines running
60 and 90 clients (requesting ﬁles of less than 10MB size
using HTTP protocol) in the 40 and 60 node network, re-
spectively. To avoid ﬂooding of network requests, clients
sleep in the 0 −60sec interval for random periods after
every random number of web requests.
4.1.3. Cloud Services
Sulaiman and Zhioua  performed extensive experi-
ments using Amazon EC2 cloud services. An Apache web
server was used to host a simple web page. socket.io
with node.js.Socket.io was installed which supported
WebSocket to help users’ browsers in using OP and us-
ing unpopular ports. For path selection, simulations were
also performed for entrance router selection algorithm and
non-entrance router selection algorithm. Several experi-
ments were conducted on a number of unpopular ports
with 1,500 circuits established per experiment and com-
promised links were detected.
Karaoglu et al.  implemented a unidirectional sce-
nario of client uploading a ﬁle to a web server. A client es-
tablished multiple socket connections for multipath trans-
missions. A 1.5MB ﬁle was uploaded through clients. To
incorporate geo-diversity, client softwares were installed in
the US and at Amazon EC2 sites in Singapore, Ireland
and North Virginia. A web server, placed at the Emulab
Utah facility, listened on multiple ports.
Biryukov et al.  performed deanonymization by
spending less than 100 USD on Amazon EC2 cloud. 50
Amazon EC2 instances were generated which captured
59,130 publication requests. Data from 120 running hid-
den services from the Tor network was collected. Collected
data was used to identify the vulnerability of Tor hidden
4.1.4. OpenFlow Enabled Network
Mendonca et al.  used OpenFlow implementation
for their proposed AnonyFlow scheme. An experimen-
tal testbed used Linux to connect the two subnetworks.
Each subnetwork was connected to two OpenFlow enabled
switches and two Net FPGA based switches. All these
OpenFlow switches were governed by a NOX controller.
iperf was used at the two client hosts with each running
for nearly 25secs.
4.1.5. Universal Composability framework
Backes et al.  provided security enhancements to
the currently used Tor network. New algorithms were
provided and setup in the universal composability (UC)
4.2. Tor Simulations
Tor simulations have been performed by (1) developing
custom simulator (2) using ExperimenTor, (3) employing
Shadow simulator, and (4) using ModelNet, as shown in
Figure 11. Figure 11 shows that 75% of the researches
(comprising of simulations) used in this study developed
their custom simulator. Only 13% used ExperimenTor.
8% and 4% of the studies used Shadow simulator and Mod-
4.2.1. Custom Simulator
Tschorsch and Scheuermann  conducted simula-
tions on ns-3 to implement Tor network with and without
N23 modiﬁcations. To replicate the onion routers environ-
ment, all onion routers were connected to a central node.
Access links of all onion routers had an 80ms delay and
100Mbps bandwidth. Sending hosts generate data at a
rate of 400kbps and Tor nodes had a maximum bandwidth
limit of 600kbps.
Doswell et al.  used the generic network simu-
lator OMNET++ to simulate mobile Tor. Wireless ac-
cess points were placed 75m apart and results were esti-
mated using linear mobility. Average throughput (kbps)
was selected as the performance metric. A 300kB webpage
was downloaded after every 2secs over the time-frame of
600secs. An artiﬁcial latency was also introduced to incor-
Edman and Syverson  implemented the multi-thread
path selection algorithm in C. Relationships between dif-
ferent ASs were borrowed from predecessor studies. RIBs
collected by University of Oregon’s RouteViews pro ject
Ngan et al.  built a discrete event simulator, in
Java, for the Tor network. 64-bit AMD Opteron 252 dual
core processors were used with 4GB RAM and operating
Figure 11: Classiﬁcation of simulations on Tor.
on Sun’s JVM and RedHat Enterprise Linux. Tor net-
work with 150 relays was simulated and all cells from ev-
ery client were simulated at every hop. Link latency was
100ms and link capacity was 500KB/s. All scenarios were
tested comprising of Tor’s original design, proposed design
and a hybrid mechanism.
Benmeziane and Badache  built their own simulator
which incorporates public communications, DNS requests,
and anonymous communications by Tor. The authors used
500 senders using 100 Tor relay nodes with 10 executions
per sender. Authors increased the number of recipients to
200. Much of the simulator details were skipped.
Li et al.  developed their own discrete event simu-
lator for Tor network. Key data structures and algorithms
of Tor were used to simulate several thousand nodes. How-
ever, authors did not perform encryption, decryption and
data transmission to avoid complexity. Moreover, simula-
tions were driven by initialization and termination events.
For a closer look, realistic values of bandwidth and up-
time were obtained from the Tor metrics portal. Eﬀective
bandwidth of relays was set to 155kBps with a 750 stan-
dard deviation. 3000 relays with millions of clients were
used for simulations.
Snader and Borisov  developed a custom ﬂow-level
simulator for the Tor network. Using the Tor metric por-
tal, bandwidth of actual Tor relays was used to simulate a
1,000 node network. 10,000 ﬂows were simulated for each
time unit of the simulator. Fair queueing was used for ﬂow
Jansen et al.  built a discrete event simulator for
the Tor network comprising of 19,400 web clients, 300 Tor
relays, 2,000 servers and 600 ﬁle sharing nodes. For ﬁle
sharing, web traﬃc comprised of 12Mbps downstream with
1.3Mbps upstream bandwidth.
Johnson et al.  built the TorPS simulator for se-
lection of Tor paths. Simulations were carried out for six
months with an adversary model containing one guard re-
lay and one exit relay having 83MBps and 16.7MBps. For
analysis of client behavior, 50,000 Monte Carlo simula-
tions were carried out spanning a period of three months.
Nowlan et al.  developed a setup for a small virtual
Tor network to estimate the performance of the proposed
modiﬁcation. Tor network comprises of three directory
authorities, three relay servers and single onion proxy. The
link delay had a mean of 50ms with a 5% path loss for the
second onion router.
Jansen et al.  performed extensive simulations for
their proposed model on both small and large-scale net-
works. Loss rate and latency have been borrowed from
Ookla and iPlane estimation services. For small scale net-
work, 50 relays and 500 clients have been conﬁgured using
50 HTTP ﬁle servers. For large scale networks, 100 re-
lays and 1,000 clients are linked with 100 HTTP servers.
Files of 320KB and 5MB are downloaded for performance
Danner et al.  carried out extensive simulations of
their proposed analytical model. However, authors do not
focus on experiments or discrete event simulations.
Wang et al.  analyzed the performance of their
proposed bridge distribution mechanism on an event-based
simulator. Aggressive blocking, conservative blocking and
event-driven based blocking of bridges were tested. The
authors also developed an analytical model for performance
Jansen and Hopper  developed a discrete event
simulator to replicate the real-world Tor network in soft-
ware running on a single machine. Performance was val-
idated against 402 node PlanetLab network. Through
HTTP client and server plugins, data was transferred through
Shadow for veriﬁcations of simulations.
Smits et al.  developed an open source imple-
mentation of the proposed mechanism. Implementation is
based on Linux version 2.6.4. Bridge distribution author-
ities needed to be reconﬁgured for distribution of keys.
Elahi et al.  simulated Tor entry guard selection
and rotation mechanism on multicore servers with each
simulation run comprising of 80,000 users. The entry
guard data was collected from real Tor network spanning
a duration of eight months.
Zhang et al.  developed a complete Tor setup con-
taining client, server and three onion routers. A probe
server and user nodes were deployed in diﬀerent network
segments. Tor code in the nodes was conﬁgured to use the
designated three relay nodes. Data from the probe server
was sent in bursts after every 0.2secs while corrupt server
sends data after every 10 −15secs.
Bauer et al.  built a toolkit for emulation of Tor
network named by ExperimenTor. The ModelNet net-
work emulation platform has been used as the baseline
approach. Scalability is one of the issues in Experimen-
Tor, owing to high resource consumption for large number
Wacek et al.  performed network experiments over
ExperimenTor for a variety of network topologies. Authors
also performed simulations on their simulator which mod-
eled a 1,524 relay network.
Gopal and Heninger  used ExperimenTor frame-
work for simulations of their proposed Torchestra approach.
ExperimenTor was setup on two physical machines work-
ing as edge node and emulator. For performance analy-
sis, small and large ﬁles were downloaded starting from
300KB. In the following stage, web and SSH traﬃc were
4.2.3. Shadow Simulator
Geddes et al.  used the Shadow simulator with real
Tor code on a simulated network. The simulated network
consisted of 160 exit relays, 240 non-exit relays, 2375 web
clients, 125 bulk clients, 150 small and medium Torperf
clients and 400 HTTP servers. Experiments consisted of
downloads of a 320KB ﬁle from random servers after ran-
dom delays (1−60secs). Bulk clients downloaded 5MB ﬁle
without any wait time. For TorPerf clients, 50KB, 1MB
and 5MB ﬁles were downloaded after every ten minutes.
Jansen et al.  performed simulations over Shadow
with a setup of 200 HTTP servers, 950 Tor web clients,
50 Tor bulk clients and 50 Tor relays. Bulk clients down-
loaded 5MB ﬁle while web clients downloaded 5KB page.
Latency of network was borrowed from the latency of Plan-
etLab nodes. Performance was estimated by varying the
load from 25 (light) to 50 (medium) to high (100) bulk
AlSabah et al.  used the ModelNet network emu-
lation platform along with practical traﬃc models for per-
formance evaluations. For small-scale experiments, 200
downloads are made of the 300KB and 5MB ﬁles by two
clients in two separate experiments. For large scale exper-
iment, 20 Tor routers are deployed with real Tor networks’
bandwidth. Each link has 80ms RTT delay. Ten clients
download 1–5MB ﬁle and 190 clients download 100–500KB
4.3. Tor’s Analysis
A number of studies limited their research works to
the analysis of current Tor network instead of simulations
and experiments of Tor. Analysis occurs in the subﬁelds
of usability of Tor, path selection mechanism, empirical
analysis and development of theoretical model. In below
lines, we present the individual studies comprising of Tor’s
4.3.1. Analytical Model
Several studies develop analytical model for analysis
of Tor network. These studies are presented in following
Anti-misbehaviour Policy Analysis: Liu and Wang 
proposed anti-misbehavior policies and analyzed it with
the original Tor architecture. No simulations or experi-
ments were conducted.
Security Analysis: Goldberg  built an analytical
model for analyzing the security of Tor’s authentication
protocol. Authors focused on analytical evaluations rather
than simulations or experiments.
Botnet Abuse Analysis: Hopper  analyzed the var-
ious possibilities for avoiding botnet abuse in the Tor net-
work. Majority schemes were discussed only, and a few
schemes were tested to verify the performance. Several
schemes were analyzed analytically.
Anonymity Model: Xin et al.  developed a theoret-
ical model to increase the anonymity of the Tor network.
They aimed to implement the proposed system on Planet-
Lab testbed, in future.
4.3.2. Empirical Analysis
A number of studies performed empirical analysis of
Tor network without performing simulations or experi-
ments. In following lines, we present the ﬁndings of these
Statistical Analysis: Wang et al.  performed em-
pirical analysis and used data available from “Tor Metric
Portal” for analysis. There were no simulations or experi-
ments performed in the research. In another study, Elices
et al.  analyzed their attack on Tor using empirical
analysis. Access logs from seven web servers were obtained
to analyze user request pattern. Moreover, Abbott et al.
 conducted a statistical evaluation by measuring the
probabilities of breaching Tor using the proposed scheme.
Robustness Analysis: Barthe et al.  analyzed the
robustness of the Tor network and proposed enhancements
in the current network. Cryptographic enhancements were
evaluated without any simulation or experimental valida-
Path Selection Protocol: Liu and Wang  presented
an improved circuit building protocol with no simulations
or experiments. Proposed algorithm was analyzed consid-
ering various aspects. In another study, Liu and Wang
 presented random walk based algorithm for Tor cir-
cuit construction. Anonymity and performance were the
key metrics evaluated in their study. However, the scope
of this study did not cover simulations or experimental
4.3.3. Usability Analysis
Usability analysis of Tor has not been carried out by a
lot of studies. However, some studies referring to usability
analysis are summarized in below lines.
Clark et al.  conducted a usability analysis by
installing various components of Tor including Vidalia,
Privoxy, Torbutton and Foxyproxy on a standard machine.
In another study, Abou-Tair et al.  presented the
usability analysis of the various anonymous service appli-
cations including Tor. Various anonymity tools were in-
stalled on a machine and usability, ease of installation and
use was analyzed.
In this section, we present the discussion and our ﬁnd-
ings of Tor network. In the ﬁrst part, we present the per-
formance metrics used to evaluate the Tor network in dif-
ferent research works. In the second part, we present our
ﬁndings of Tor research works referred in this study. In
the last part, we show our ﬁndings for open research areas
in the ﬁeld of Tor network which may be used for future
5.1. Tor Performance Metrics
Analyzing the performance metrics is a crucial task
for future research, analysis, simulations and experiments
in the Tor network. Table 10 presents the performance
metrics of Tor used in various studies. No clear patterns
were observed, so, authors described the metrics used in
individual studies. A brief overview of the table shows
that throughput (bandwidth) and latency are the most fre-
quently used metrics. However, every research formalized
its own performance metric based upon the requirement
of the experiment.
5.2. Survey Findings
In this section, we summarize our ﬁndings for the onion
router by comparing all studies with a deep focus over the
key concepts and ideas used in diﬀerent research works.
We divide our research evaluations in three subcategories
considering (1) research areas, (2) research platforms, and
(3) performance metrics.
5.2.1. Research Areas
The majority of Tor research (nearly 55%) covering
anonymity is focused over deanonymization of Tor net-
work. Around 20% studies are related to the path selection
mechanism. Only 25% research studies are on performance
analysis and improvement mechanism of Tor network. Ac-
cording to Dingledine (the co-founder of Tor pro ject), ma-
jority research works focus their attention on the breaching
In the deanonymization track, 35% of the studies de-
sign deanonymization attacks for Tor while 21% deanonymize
Tor using traﬃc analysis. 16% focus on improvements to
bypass deanonymization while 14% study ﬁngerprinting
mechanisms to identify Tor traﬃc on the Internet. Only
9% identify hidden services while 2% focus on anonymity
mechanisms without using Tor.
All deanonymization related Tor studies have exploited
its inherent weaknesses. Compromised relays are the most
exploited weaknesses followed by traﬃc interception and
protocol messages. Very few studies focus on the com-
promised autonomous systems, browsers, servers, decoy
traﬃc, and ﬂag cheating.
In the path selection track, 87% of the studies focus
on the design of new path selection algorithms and 13%
research works analyze currently developed algorithms.
Our analysis shows that anonymity and performance
(bandwidth and latency) are the most important param-
eters used in the design and analysis of path selection al-
gorithms. Relays have been incorporated in the design of
path selection algorithms covering both location and ca-
pacity of relays. Other parameters include autonomous
systems, hops, multi-path mechanism and load.
Performance Analysis and Architectural Improvements:
In the performance analysis and improvement track,
32% of the research works cover analysis and 27% of stud-
ies focus on performance improvement mechanisms. 29%
of the studies provide general analysis of Tor covering us-
ability and sociability issues. 10% of the research works
focus on modeling of the Tor network while 2% address
Analysis of various research studies show that perfor-
mance (latency and bandwidth), relay selection, and anonymity
are the most used parameters. Other studies also pay at-
tention to queues, QoS, protocol messages and traﬃc shap-
5.2.2. Research Platforms
An interesting feature revealed in analysis is the fact
that 60% of the research works were conducted by perform-
ing real-world experiments on the Tor network. Although
special measures were taken to protect the identity of users
but majority research works failed to analyze legal or eth-
ical requirements of capturing user data and performing
experiments by developing attacks in real network. Only
27% of studies developed their own simulator and 13%
conducted analysis without experiments or simulations.
Our survey shows that 86% of the research works devel-
oped their own testbed for experiments. The majority
of studies deployed 1-2 clients with 1-3 servers for exper-
iments. Research works covering relays used 1-3 relays.
However, some research works increased the number of re-
lays by using virtual machines and PlanetLab. A limited
number of studies used cloud-based setups.
Interestingly, 75% of the research works developed their
own simulator without any common parameters used for
Tor network. ns-3, OMNET, C and Java were used for the
development of custom simulators. 13% of the research
works used ExperimenTor. Our research shows that Ex-
perimenTor is the most common toolkit used by majority
of the research. 8% and 3% of the studies used Shadow
simulator and ModelNet, respectively.
5.2.3. Performance Metrics
We considered the performance metrics used in various
works. Analysis shows that no hard and fast rule exists
for use of performance metrics. Every study developed its
own metrics to measure performance, anonymity and QoS.
Moreover, no baseline techniques exist for the comparison
5.3. Open Research Areas
Our survey shows that majority of the research works
are concentrated in a few domains. However, a number
of major challenges exist owing to the peer-to-peer nature
of Tor. A number of key areas have also been identiﬁed
by the Tor team. We identiﬁed the following areas which
require further research.
1. Data Estimation: Estimation of key network statis-
tics is the most critical task in the Tor network be-
cause it is a peer-to-peer network. No one can see
the entire traﬃc so it is not possible to estimate the
size of Tor network. Some of the statistics requiring
attention are as follows:
Table 10: Performance metrics used in various research works.
Domain Performance Metrics Research Works
Throughput; Bandwidth; Packet rate; Bit rate;
Mendonca et al. , Zhang et al. , Pries et al. ,
Jin and Wang , Marks et al. , Panchenko et al.
, Chen and Pasquale , Karaoglu et al. , Li
et al. , Panchenko et al. , Snader and Borisov
, Houmansadr et al., Pries et al. , Tschorsch
and Scheuermann , Andersson and Panchenko ,
Doswell et al. , Jansen et al. , Moghaddam
et al., Weinberg et al. , Jansen et al. , Ehlert
, Barbera et al. , Hopper , Nowlan et al.
, Ngan et al. , Panchenko et al. , Wang et
al. , Jansen et al. , Tang and Goldberg ,
AlSabah et al., Wack et al. , Geddes et al. ,
AlSabah et al. , Jansen and Hopper et al. ,
Jansen et al. 
Latency; Webpage loading time; Round trip time;
Download Time; Router latency; Circuit setup
duration; Boot strap duration; Time to ﬁrst byte;
Time to last byte; Ping reply delay; Per hop la-
tency; SYN and SYN ACK diﬀerence; Delay per
cell; Jitter; Inter packets delay distribution
Mendonca et al. , Overlier and Syverson , Loesing
et al. , Chan-Tin et al. , Herzberg et al. ,
Murdoch and Danezis , Zhang et al. , Akhoondi et
al. , Panchenko et al. , Li et al. , Dhungel and
Steiner , Andersson and Panchenko , Doswell et
al. , AlSabah et al., Moghaddam et al., Weinberg et
al. , Evans et al. , Wang et al. , Jansen et
al. , Ehlert , Hopper , Winter and Lind-
skog , Edmundson et al. , Ngan et al. ,
Panchenko et al. , Lenhard et al. , Wack et al.
, Geddes et al. , AlSabah et al. , Jansen and
Hopper et al. , Snader and Borisov , Jansen et
al. , Chen et al. , Smits et al. , Gopal and
Heninger , Panchenko et al. , Panchenko et al.
True Positive; True Negative; False Positive; False
Negative; Region of Convergence; Recognition
rate; Mis-recognition rate; Accuracy; Recall; Pre-
Chakravarti et al. , Barker et al. , Chan-Tin et al.
, Akhoondi et al. , Danner et al. , Gilad and
Herzberg , Panchenko et al. , Song et al., Wang
and Goldberg , Wang and Goldberg , Elices et al.
, Bai et al. , AlSabah et al., Wagner et al. 
Timing Attack correlation Overlier and Syverson , Zhang et al. , Pries et al.
, Wang et al. , Houmansadr et al., Murdoch and
Danezis , Song et al., Panchenko et al. 
Compromised relays; Compromised circuits;
Compromised Streams; Time for ﬁrst compro-
mised stream; Failure rate; Compromise time;
Compromised links; Compromised Tunnels; De-
tection rate; Compromised Clients; Compromised
router bandwidth; Compromise probability; Con-
gestion attack time
Overlier and Syverson , Sulaiman and Zhioua ,
Chen and Pasquale , Li et al. , Li et al. ,
Bauer et al. , Johnson et al. , Evans et al. ,
Danner et al. , Chakravarty et al. , Snader and
Borisov , Panchenko et al. , Elahi et al. ,
Abbott et al. , Bauer et al. , Wang et al. 
Packet Sizes; Probability diﬀerence plots; Energy
plots; Recipient probabilities; Queued messages
length; Anonymity vs performance; Router band-
width; HTTP content distribution; Tor servers;
Tor traﬃc; Generated Paths; Client resource
usuage; Tor load per circuit; Node Connection
pattern; IP TTL diﬀerence; Service, browser, ﬁle
format usuage; Tor location usuage; Boot strap
time; Exit traﬃc stats; Tor bridges statistics; hid-
den service descriptor request rate; Botnet decay
rate; Tor overhead; Router statistics
Barker et al. , Loesing et al. , Benmeziane et al.,
Jin and Wang , Zhang et al. , Liu and Wang ,
Liu and Wang , Dhungel and Steiner , Chaa-
bane et al. , Mulazzani et al. , Moghaddam et
al., Edman and Syverson , Barbera et al. , Hop-
per , Winter and Lindskog , Huber et al. ,
Blond et al. , Lenhard et al. , McCoy et al.
, Wang et al. , Biryukov et al. , Loesing et
al. , Chen et al. , Panchenko et al. , Elahi
et al. , Marks et al. 
Empirical Evaluations: Usability analysis, secu-
rity model analysis, general discussion, proposed
Clark et al. , Abou-Tair et al. , Goldberg ,
Kuhn et al., Barthe et al. , Gros et al. 
•Number of clients in the network: Peer-to-peer
networks make it impossible to estimate the to-
tal traﬃc statistics because no user can see the
•Capabilities of relays: There is limited infor-
mation available about the relays which are the
most crucial parameters in path selection. In-
corporation of relay capabilities into anonymity
of Tor and performance model is a key research
area as done in a number of studies.
•Performance of the network: Estimation of net-
work performance at any given time is a crucial
task. Owing to the P2P nature, only health of
relays is known to the Tor administration. How
is the network performing at any given instant?
is still a crucial task.
•Number of clients connecting via bridges: Tor
authorities provide secret relay addresses to clients
who can’t access Tor due to blockage of re-
lays in their location. However, very little is
known about the quantity of clients connecting
through bridges and their traﬃc statistics.
•Exit network traﬃc: Signiﬁcant research is re-
quired about the exit network traﬃc. All clients
pass their data through relays and very little is
known about the statistics of traﬃc exiting exit
2. Analysis: Deep analysis of the current Tor network is
required. Analysis may be based upon an extension
of previous research into path length, anonymity, la-
tency, etc. Analysis of the optimal performance pa-
rameters is required.
3. Measurement and Attack tools: Development of novel
attack methodologies to identify the shortcomings
of the current Tor network. Tor has no automatic
mechanism to identify anomalies and assess the health
of the network. Attack tools should be developed
which should prevent attacks occurring from com-
promised relays and servers. Comprised relays are
vulnerable to botnet based attacks comprising of DDOS
attacks, ﬁngerprinting attacks etc. Despite large amount
of research in botnet attacks, it is still open to re-
search which would make Tor a more stable and se-
4. Defenses against Attacks: Develop novel defense method-
ologies to counter attacks on the Tor network. Al-
though majority research works have focused on the
development of novel attack methodologies, very lit-
tle is known about viable counter-measures. Our sur-
vey shows that relays are mostly vulnerable because
they can be deployed by any eavesdropper. Counter-
measures against congestion attacks, latency mea-
suring attacks, throughput measuring attacks, etc.
can help in the improvement of Tor.
This paper deals with the survey, classiﬁcation, quan-
tiﬁcation and comparative analysis of various research works
covering Tor network. To the author’s best knowledge, no
other survey/research has performed such a deep and thor-
ough analysis of Tor studies. Our study shows that Tor re-
search areas can be broadly classiﬁed into (1) deanonymiza-
tion, (2) path selection, (3) analysis and performance im-
provements. More than half studies carried out address
‘deanonymization’ with major subdivisions into deanonymiza-
tion ‘attacks’ and ‘traﬃc analysis’ attacks. In the ‘path se-
lection’ area, more than 85% of the studies have focused on
the development of new algorithms. In the ‘analysis and
performance improvement’ area, the majority of studies
are a mixed bag, followed by analysis, followed by perfor-
mance improvement studies. Our analysis of Tor platforms
shows that 60% of studies performed experiments while
27% performed simulations. Among experiments, 86% of
the studies deployed private testbeds. Among simulations,
75% developed their own simulators. Analysis of parame-
ters (used in various studies) shows that their is no little
consistency across various studies. However, a majority of
the studies used variations of throughput and latency for
 Russell Brandom. FBI agents tracked Harvard bomb
threats despite Tor. THE VERGE. Available at
tracked-harvard-bomb-threats-across-tor, December 18, 2013.
 Ilya Khrennikov. Russians Find Ways to By-
pass Latest Web Ban. Bloomberg. Available at
blocking, February 2, 2016.
 Alex Hern. US defence department funded Carnegie Mel-
lon research to break Tor. The guardian. Available at
tor, 25 February 2016.
 Darien Graham-Smith. Extreme online se-
curity measures to protect your digital pri-
vacy a guide. The guardian. Available at
security-measures-digital-privacy-guide, 3 July 2016.
 Dave Neal. Mozilla will have to wait to ﬁnd out
how the FBI cracked Tor. The Inquirer. Available at
wants-to-know-how-the-fbi-cracked-tor, 18 May 2016.
 Mashael AlSabah and Ian Goldberg. Performance and security
improvements for tor: A survey. ACM Computing Surveys
(CSUR), 49(2):32, 2016.
 R. Koch, M. Golling, and G. D. Rodosek. How Anonymous
Is the Tor Network? A Long-Term Black-Box Investigation.
Computer, 49(3):42–49, Mar 2016.
 Mashael AlSabah and Ian Goldberg. Performance and Security
Improvements for Tor: A Survey. IACR Cryptology ePrint
Archive, 2015:235, 2015.
 Bernd Conrad and Fatemeh Shirazi. A Survey on Tor and
I2P. Proceedings of 9th International Conference on Internet
Monitoring and Protection (ICIMP), page 22, 2014.
 Rolf Jagerman, Wendo Sabee, Laurens Versluis, Martijn
de Vos, and Johan Pouwelse. The ﬁfteen year struggle of
decentralizing privacy-enhancing technology. arXiv preprint
 Jian Ren and Jie Wu. Survey on anonymous communications
in computer networks. Computer Communications, 33(4):420–
 Peter C Johnson and Apu Kapadia. From Chaum
to Tor and Beyond: A Survey of Anonymous Rout-
ing Systems. Dartmouth Col lege (Technical Re-
port), Available at http://www.cs.dartmouth.edu/ cc-
 David Goldschlag, Michael Reed, and Paul Syverson. Onion
routing. Communications of the ACM, 42(2):39–41, 1999.
 TorMETRICS. Relays and bridges in the network. Available
at https://metrics.torproject.org/, July 13, 2016.
 Michael G Reed, Paul F Syverson, and David M Goldschlag.
Proxies for anonymous routing. In Annual Computer Security
Applications Conference, pages 95–104. IEEE, 1996.
 David M Goldschlag, Michael G Reed, and Paul F Syverson.
Hiding routing information. In International Workshop on
Information Hiding, pages 137–150. Springer, 1996.
 Michael G Reed, Paul F Syverson, and David M Goldschlag.
Anonymous connections and onion routing. IEEE Journal on
Selected areas in Communications, 16(4):482–494, 1998.
 Paul Syverson, Gene Tsudik, Michael Reed, and Carl
Landwehr. Towards an analysis of onion routing security.
In Designing Privacy Enhancing Technologies, pages 96–114.
 Emmanuel Bresson, Olivier Chevassut, David Pointcheval, and
Jean-Jacques Quisquater. Provably authenticated group diﬃe-
hellman key exchange. In Proceedings of the 8th ACM confer-
ence on Computer and Communications Security, pages 255–
264. ACM, 2001.
 Ryan Pries, Wei Yu, Xinwen Fu, and Wei Zhao. A new re-
play attack against anonymous communication networks. In
International Conference on Communications (ICC), pages
1578–1582. IEEE, 2008.
 Marco Valerio Barbera, Vasileios P Kemerlis, Vasilis Pappas,
and Angelos D Keromytis. Cellﬂood: Attacking Tor onion
routers on the cheap. In European Symposium on Research
in Computer Security - ESORICS, pages 664–681. Springer,
 Tao Wang and Ian Goldberg. Improved website ﬁngerprinting
on Tor. In Proceedings of the Workshop on Privacy in the
Electronic Society (WPES), pages 201–212. ACM, 2013.
 Adam Gordon and Steven Hernandez. The Oﬃcial (ISC) 2
Guide to the SSCP CBK. John Wiley & Sons, 2016.
 SoftEther VPN Project. University of Tsukuba.
https://www.softether.org, Last Accessed: July 2016.
 JanusVM: An Internet Privacy Appliance.
http://janusvm.peertech.org, Last Accessed: July 2016.
 proXPN. https://www.proxpn.com, Last Accessed: July 2016.
 USAIP. https://usaip.eu, Last Accessed: July 2016.
 VPNreactor. https://www.vpnreactor.com/, Last Accessed:
 Xero Networks AG and Steve Topletz. xB Browser.
http://xerobank.com/, Last Accessed: 2009.
 AnchorFree Inc. Hotspot Shield.
http://www.hotspotshield.com/, Last Accessed: July 2016.
 Advanced Onion Router. https://sourceforge.net/projects/advtor/,
Last Accessed: July 2016.
 SecurityKISS. https://www.securitykiss.com/, Last Accessed:
 UltraReach. UltraSurf. http://ultrasurf.us/, Last Accessed:
 CyberGhost S.R.L. CyberGhost VPN.
http://www.cyberghostvpn.com/, Last Accessed: July 2016.
 Dynamic Internet Technology Inc. (DIT). Freegate. http://dit-
inc.us/freegate.html, Last Accessed: July 2016.
 Tails. https://tails.boum.org/, Last Accessed: July 2016.
 Privatix. http://www.mandalka.name/privatix/, Last Ac-
cessed: July 2016.
 Surfboard Holdings B.V. Ixquick. https://www.ixquick.com/,
Last Accessed: July 2016.
 Inc. DuckDuckGo. DuckDuckGo. https://duckduckgo.com/,
Last Accessed: July 2016.
 Anonymous Email. https://anonymousemail.me/, Last Ac-
cessed July 2016.
 Safe-mail. http://www.safe-mail.net/, Last Accessed: July
 Hushmail. https://www.hushmail.com/, Last Accessed: July
 10minutemail. https://10minutemail.com, Last Accessed:
 YOPmail. http://www.yopmail.com, Last Accessed: July
 Daniel Arp, Fabian Yamaguchi, and Konrad Rieck. Torben:
Deanonymizing Tor communication using web page markers.
Technical report, IFI-TB-2014-01, University of G¨ottingen,
 Lasse Overlier and Paul Syverson. Locating hidden servers.
In Symposium on Security and Privacy, pages 15–pp. IEEE,
 Juan A Elices, Fernando Perez-Gonzalez, and Carmela Tron-
coso. Fingerprinting Tor’s hidden service log ﬁles using a tim-
ing channel. In International Workshop on Information Foren-
sics and Security (WIFS), pages 1–6. IEEE, 2011.
 Lu Zhang, Junzhou Luo, Ming Yang, and Gaofeng He.
Application-level attack against Tor’s hidden service. In In-
ternational Conference on Pervasive Computing and Applica-
tions (ICPCA), pages 509–516. IEEE, 2011.
 Alex Biryukov, Ivan Pustogarov, and R Weinmann. Trawl-
ing for Tor hidden services: Detection, measurement,
deanonymization. In Symposium on Security and Privacy,
pages 80–94. IEEE, 2013.
 Xuefeng Bai, Yong Zhang, and Xiamu Niu. Traﬃc identiﬁca-
tion of Tor and web-mix. In International Conference on In-
telligent Systems Design and Applications (ISDA), volume 1,
pages 548–551. IEEE, 2008.
 John Barker, Peter Hannay, and Patryk Szewczyk. Using
Traﬃc Analysis to Identify the Second Generation Onion
Router. In International Conference on Embedded and Ubiq-
uitous Computing (EUC), pages 72–78. IEEE, 2011.
 Mashael AlSabah, Kevin Bauer, and Ian Goldberg. Enhancing
Tor’s performance using real-time traﬃc classiﬁcation. In Pro-
ceedings of the conference on Computer and Communications
Security, pages 73–84. ACM, 2012.
 Amir Houmansadr, Chad Brubaker, and Vitaly Shmatikov.
The parrot is dead: Observing unobservable network commu-
nications. In Symposium on Security and Privacy, pages 65–
79. IEEE, 2013.
 Sambuddho Chakravarty, Angelos Stavrou, and Angelos D
Keromytis. Identifying proxy nodes in a Tor anonymization
circuit. In International Conference on Signal Image Tech-
nology and Internet Based Systems (SITIS), pages 633–639.
 Philipp Winter and Stefan Lindskog. How the great ﬁrewall
of China is blocking Tor. Free and Open Communications on
the Internet (FOCI), 2012.
 Muhammad Aliyu Sulaiman and Sami Zhioua. Attacking Tor
through Unpopular Ports. In International Conference on
Distributed Computing Systems Workshops (ICDCSW), pages
33–38. IEEE, 2013.
 Eric Chan-Tin, Jiyoung Shin, and Jiangmin Yu. Revisiting
Circuit Clogging Attacks on Tor. In International Conference
on Availability, Reliability and Security (ARES), pages 131–
140. IEEE, 2013.
 Xiaogang Wang, Junzhou Luo, Ming Yang, and Zhen Ling.
A novel ﬂow multiplication attack against Tor. In Interna-
tional Conference on Computer Supported Cooperative Work
in Design (CSCWD), pages 686–691. IEEE, 2009.
 Cynthia Wagner, Gerard Wagener, Radu State, Alexandre Du-
launoy, and Thomas Engel. Breaking Tor anonymity with
game theory and data mining. Concurrency and Computa-
tion: Practice and Experience, 24(10):1052–1065, 2012.
 Souad Benmeziane and Nadjib Badache. Tor network lim-
its. Technical report, CERIST Digital Library, Availablle at
 Rob Jansen, Florian Tschorsch, Aaron Johnson, and Bj¨orn
Scheuermann. The Sniper Attack: Anonymously Deanonymiz-
ing and Disabling the Tor Network. In Network and Dis-
tributed Systems Security Symposium (NDSS), San Diego,
CA, USA, 2014.
 Timothy G Abbott, Katherine J Lai, Michael R Lieberman,
and Eric C Price. Browser-based attacks on Tor. In Privacy
Enhancing Technologies, pages 184–199. Springer, 2007.
 Nathan S Evans, Roger Dingledine, and Christian Grothoﬀ.
A Practical Congestion Attack on Tor Using Long Paths. In
USENIX Security Symposium, pages 33–50, 2009.
 Kevin Bauer, Damon McCoy, Dirk Grunwald, Tadayoshi
Kohno, and Douglas Sicker. Low-resource routing attacks
against Tor. In Proceedings of the workshop on Privacy in
electronic society, pages 11–20. ACM, 2007.
 Matthew Edman and Paul Syverson. AS-awareness in Tor
path selection. In Proceedings of the conference on Computer
and Communications Security, pages 380–389. ACM, 2009.
 Stevens Le Blond, Pere Manils, Chaabane Abdelberi, Mo-
hamed Ali Dali Kaafar, Claude Castelluccia, Arnaud Legout,
and Walid Dabbous. One bad apple spoils the bunch: exploit-
ing P2P applications to trace and proﬁle Tor users. arXiv
preprint arXiv:1103.1518, 2011.
 John Geddes, Rob Jansen, and Nicholas Hopper. How low can
you go: Balancing performance with anonymity in Tor. In
International Symposium on Privacy Enhancing Technologies
(PETS), pages 164–184. Springer, 2013.
 Sambuddho Chakravarty, Georgios Portokalidis, Michalis
Polychronakis, and Angelos D Keromytis. Detecting traﬃc
snooping in Tor using decoys. In Recent Advances in Intru-
sion Detection (RAID), pages 222–241. Springer, 2011.
 Aaron Johnson, Chris Wacek, Rob Jansen, Micah Sherr, and
Paul Syverson. Users get routed: Traﬃc correlation on Tor
by realistic adversaries. In Proceedings of the conference on
Computer & Communications Security (CCS), pages 337–348.
 Steven J Murdoch and George Danezis. Low-cost traﬃc anal-
ysis of Tor. In Symposium on Security and Privacy, pages
183–195. IEEE, 2005.
 Sambuddho Chakravarty, Marco V Barbera, Georgios Por-
tokalidis, Michalis Polychronakis, and Angelos D Keromytis.
On the Eﬀectiveness of Traﬃc Analysis Against Anonymity
Networks Using Flow Records. In International Conference
on Passive and Active Measurement (PAM), pages 247–257.
 Jia Zhang, Haixin Duan, and Jianping Wu. A Novel Method to
Prevent Traﬃc Analysis in Low-Latency Anonymous Commu-
nication Systems. In Proceedings of the International Confer-
ence on Computer and Electrical Engineering, pages 906–911.
 Ming Song, Gang Xiong, Zhenzhen Li, Junrui Peng, and
Li Guo. A de-anonymize attack method based on traﬃc
analysis. In International ICST Conference on Communica-
tions and Networking in China (CHINACOM), pages 455–460.
 Andriy Panchenko, Lukas Niessen, Andreas Zinnen, and
Thomas Engel. Website ﬁngerprinting in onion routing based
anonymization networks. In Proceedings of the annual work-
shop on Privacy in the Electronic Society (WPES), pages 103–
114. ACM, 2011.
 Jing Jin and Xinyuan Wang. On the eﬀectiveness of low la-
tency anonymous network in the presence of timing attack. In
International Conference on Dependable Systems & Networks
(DSN), pages 429–438. IEEE, 2009.
 Yossi Gilad and Amir Herzberg. Spying in the dark: TCP and
Tor traﬃc analysis. In International Symposium on Privacy
Enhancing Technologies, pages 100–119. Springer, 2012.
 Stjepan Groˇs, Marko Salki´c, and Ivan ˇ
Sipka. Protecting TOR
exit nodes from abuse. In Proceedings of the International
Convention on Information and Communication Technology,
Electronics and Microelectronics (MIPRO), pages 1246–1249.
 Philipp Winter, Richard K¨ower, Martin Mulazzani, Markus
Huber, Sebastian Schrittwieser, Stefan Lindskog, and Edgar
Weippl. Spoiled Onions: Exposing Malicious Tor Exit Relays.
In International Privacy Enhancing Technologies Symposium
(PETS), pages 304–331. Springer, 2014.
 Liu Xin and Wang Neng. Design Improvement for Tor Against
Low-Cost Traﬃc Attack and Low-Resource Routing Attack.
In International Conference on Communications and Mobile
Computing (CMC), volume 3, pages 549–554. IEEE, 2009.
 Michael Backes, Ian Goldberg, Aniket Kate, and Esfandiar
Mohammadi. Provably secure and practical onion routing.
In Computer Security Foundations Symposium (CSF), pages
369–385. IEEE, 2012.
 Daniel Marks, Florian Tschorsch, and Bj¨orn Scheuermann.
Unleashing Tor, BitTorrent & co.: How to relieve TCP de-
ﬁciencies in overlays. In Conference on Local Computer Net-
works (LCN), pages 320–323. IEEE, 2010.
 Michael F Nowlan, David Wolinsky, and Bryan Ford. Reducing
latency in Tor circuits with unordered delivery. In USENIX
Workshop on Free and Open Communications on the Internet
 Norman Danner, Sam Defabbia-Kane, Danny Krizanc, and
Marc Liberatore. Eﬀectiveness and detection of denial-of-
service attacks in Tor. ACM Transactions on Information
and System Security (TISSEC), 15(3):11, 2012.
 Amir Herzberg, Ely Porat, Nir Soﬀer, and Erez Waisbard.
Camouﬂaged Private Communication. In Privacy, security,
risk and trust (PASSAT), international conference on social
computing (socialcom), pages 1159–1162. IEEE, 2011.
 Marc Mendonca, Srini Seetharaman, and Katia Obraczka.
A ﬂexible in-network IP anonymization service. In Interna-
tional Conference on Communications (ICC), pages 6651–
6656. IEEE, 2012.
 Steven J Murdoch. Hot or not: Revealing hidden services
by their clock skew. In Proceedings of the conference on
Computer and Communications Security (CCS), pages 27–36.
 Sambuddho Chakravarty, Angelos Stavrou, and Angelos D
Keromytis. Linkwidth: a method to measure link capacity
and available bandwidth using single-end probes. Computer
Science Department Technical Report (CUCS Tech Report)
CUCS-002, Columbia University, 2008.
 Se Eun Oh, Shuai Li, and Nicholas Hopper. Fingerprinting
keywords in search queries over tor. Proceedings on Privacy
Enhancing Technologies, 2017(4):251–270, 2017.
 Joan Daemen and Vincent Rijmen. The design of Rijndael:
AES-the advanced encryption standard. Springer Science &
Business Media, 2013.
 NIST-FIPS Standard. Announcing the advanced encryption
standard (aes). Federal Information Processing Standards
Publication, 197:1–51, 2001.
 Daniel Marks, Florian Tschorsch, Bjrn Scheuermann, Daniel
Marks, Florian Tschorsch, and Bjrn Scheuermann. Unleash-
ing tor, bittorrent and co.: How to relieve tcp deﬁciencies in
overlays (extended version). Heinrich Heine University, Dus-
seldorf, Germany, 2010.
 Nikita Borisov, George Danezis, Prateek Mittal, and Parisa
Tabriz. Denial of service or denial of security? In Proceedings
of the 14th ACM conference on Computer and communica-
tions security, pages 92–102. ACM, 2007.
 Roger Dingledine, Nick Mathewson, and Paul Syverson. Tor:
The second-generation onion router. Technical report, DTIC
 Masoud Akhoondi, Curtis Yu, and Harsha V Madhyastha.
LASTor: A low-latency AS-aware Tor client. In Symposium
on Security and Privacy (SP), pages 476–490. IEEE, 2012.
 Fallon Chen and Joseph Pasquale. Toward improving path
selection in Tor. In Global Telecommunications Conference
(GLOBECOM), pages 1–6. IEEE, 2010.
 Hasan T Karaoglu, Mehmet Burak Akgun, Mehmet Hadi
Gunes, and Murat Yuksel. Multi Path Considerations for
Anonymized Routing: Challenges and Opportunities. In Inter-
national Conference on New Technologies, Mobility and Secu-
rity (NTMS), pages 1–5. IEEE, 2012.
 Andriy Panchenko, Fabian Lanze, and Thomas Engel. Im-
proving performance and anonymity in the tor network. In
Performance Computing and Communications Conference
(IPCCC), 2012 IEEE 31st International, pages 1–10. IEEE,
 Chenglong Li, Yibo Xue, Longtao He, and Lidong Wang.
TMT: A new Tunable Mechanism of Tor based on the path
length. In International Conference on Cloud Computing and
Intelligent Systems (CCIS), volume 2, pages 661–665. IEEE,
 Andriy Panchenko, Lexi Pimenidis, and Johannes Renner. Per-
formance analysis of anonymous communication channels pro-
vided by Tor. In International Conference on Availability,
Reliability and Security (ARES), pages 221–228. IEEE, 2008.
 Xin Liu and Neng Wang. RandomWalk-Based Tor Circuit
Building Protocol. In International Conference on Computa-
tional Intelligence and Security (CIS), volume 2, pages 335–
340. IEEE, 2009.
 Xin Liu and Neng Wang. An Improved Tor Circuit-Building
Protocol. In International Joint Conference on Artiﬁcial In-
telligence (JCAI), pages 671–675. IEEE, 2009.
 Robin Snader and Nikita Borisov. Improving security and per-
formance in the Tor network through tunable path selection.
Transactions on Dependable and Secure Computing, 8(5):728–
 Chenglong Li, Yibo Xue, Yingfei Dong, and Dongsheng Wang.
Relay recommendation system (RRS) and selective anonymity
for Tor. In Global Communications Conference (GLOBE-
COM), pages 833–838. IEEE, 2012.
 Can Tang and Ian Goldberg. An improved algorithm for Tor
circuit scheduling. In Proceedings of the conference on Com-
puter and Communications Security (CCS), pages 329–339.
 Tao Wang, Kevin Bauer, Clara Forero, and Ian Goldberg.
Congestion-aware path selection for Tor. In Financial Cryp-
tography and Data Security, pages 98–113. Springer, 2012.
 Robin Snader and Nikita Borisov. A Tune-up for Tor: Improv-
ing Security and Performance in the Tor Network. In Network
and Distributed System Symposium (NDSS), volume 8, page
 Tariq Elahi, Kevin Bauer, Mashael AlSabah, Roger Dingle-
dine, and Ian Goldberg. Changing of the guards: A framework
for understanding and improving entry guard selection in Tor.
In Proceedings of the workshop on Privacy in the electronic
society, pages 43–54. ACM, 2012.
 Kevin Bauer, Dirk Grunwald, and Douglas Sicker. Pre-
dicting Tor path compromise by exit port. In Interna-
tional Performance Computing and Communications Confer-
ence (IPCCC), pages 384–387. IEEE, 2009.
 Chris Wacek, Henry Tan, Kevin S Bauer, and Micah Sherr. An
Empirical Evaluation of Relay Selection in Tor. In Network
and Distributed System Symposium (NDSS), 2013.
 Aaron Johnson, Rob Jansen, Aaron D Jaggard, Joan Feigen-
baum, and Paul Syverson. Avoiding the man on the wire:
Improving tor’s security with trust-aware path selection. In
the Proceedings of the Network and Distributed Security Sym-
posium - NDSS, 2017, 2017.
 Xin Liu and Neng Wang. Anti-misbehavior System for Tor
Network. In International Joint Conference on INC, IMS and
IDC (NCM), pages 257–261. IEEE, 2009.
 Roger Dingledine, Nick Mathewson, and Paul Syverson. De-
ploying low-latency anonymity: Design challenges and social
factors. IEEE Security & Privacy, 5(5):83–87, 2007.
 Dhiah el Diehn I Abou-Tair, Lexi Pimenidis, Jens Schomburg,
and Benedikt Westermann. Usability inspection of anonymity
networks. In Proceedings of the World Congress on Privacy,
Security, Trust and the Management of e-Business, pages 100–
109. IEEE, 2009.
 Jeremy Clark, Paul C Van Oorschot, and Carlisle Adams. Us-
ability of anonymous web browsing: an examination of Tor
interfaces and deployability. In Proceedings of the symposium
on Usable privacy and security, pages 41–51. ACM, 2007.
 Anne Edmundson, Simpson AKornfeld, Joshua A Kroll, and
Edward W Felten. Security Audit of Safeplug “Tor in a Box”.
In Workshop on Free and Open Communications on the In-
ternet (FOCI). USENIX Association.
 Gilles Barthe, Alejandro Hevia, Zhengqin Luo, Tamara Rezk,
and Bogdan Warinschi. Robustness Guarantees for Anonymity.
In Computer Security Foundations Symposium (CSF), pages
91–106. IEEE, 2010.
 Martin Mulazzani, Markus Huber, and Edgar R Weippl.
Anonymity and monitoring: how to monitor the infrastruc-
ture of an anonymity system. IEEE Transactions on Sys-
tems, Man, and Cybernetics, Part C: Applications and Re-
views, 40(5):539–546, 2010.
 Markus Huber, Martin Mulazzani, and Edgar Weippl. Tor
HTTP usage and information leakage. In Communications
and Multimedia Security, pages 245–255. Springer, 2010.
 Damon McCoy, Kevin Bauer, Dirk Grunwald, Tadayoshi
Kohno, and Douglas Sicker. Shining light in dark places: Un-
derstanding the Tor network. In Privacy Enhancing Technolo-
gies, pages 63–76. Springer, 2008.
 Karsten Loesing, Steven J Murdoch, and Roger Dingledine. A
case study on measuring statistical data in the Tor anonymity
network. In Financial Cryptography and Data Security, pages
203–215. Springer, 2010.
 Yao Chen, Radu Sion, and Bogdan Carbunar. Xpay: Practi-
cal anonymous payments for Tor routing and other networked
services. In Proceedings of the workshop on Privacy in the
electronic society, pages 41–50. ACM, 2009.
 Rob Jansen, Kevin S Bauer, Nicholas Hopper, and Roger Din-
gledine. Methodically modeling the tor network. In Work-
shop on Cyber Security Experimentation and Test (CSET).
 Rob Jansen and Nicholas Hooper. Shadow: Running Tor in a
box for accurate and eﬃcient experimentation. Technical re-
port, Minnesota University, Department of Computer Science
and Engineering (No. TR-11-020), 2011.
 Kevin S Bauer, Micah Sherr, and Dirk Grunwald. Experi-
mentor: A Testbed for Safe and Realistic Tor Experimenta-
tion. In Workshop on Cyber Security Experimentation and
Test (CSET). USENIX, 2011.
 Prithula Dhungel, Moritz Steiner, Ivinko Rimac, Volker Hilt,
and Keith W Ross. Waiting for anonymity: Understanding
delays in the Tor overlay. In International Conference on Peer-
to-Peer Computing (P2P), pages 1–4. IEEE, 2010.
 Karsten Loesing, Werner Sandmann, Christian Wilms, and
Guido Wirtz. Performance measurements and statistics of Tor
hidden services. In International Symposium on Applications
and the Internet (SAINT), pages 1–7. IEEE, 2008.
 Mathias Ehlert. I2P Usability vs. Tor Usability A Bandwidth
and Latency Comparison. In Seminar Report, Humboldt Uni-
versity of Berlin, 2011.
 Ryan Pries, Wei Yu, Steve Graham, and Xinwen Fu. On per-
formance bottleneck of anonymous communication networks.
In International Symposium on Paral lel and Distributed Pro-
cessing (IPDPS), pages 1–11. IEEE, 2008.
 Xiao Wang, Jinqiao Shi, Binxing Fang, and Li Guo. An empiri-
cal analysis of family in the Tor network. In International Con-
ference on Communications (ICC), pages 1995–2000. IEEE,
 Florian Tschorsch and Bj¨orn Scheuermann. Tor is unfair –
And what to do about it. In Conference on Local Computer
Networks (LCN), pages 432–440. IEEE, 2011.
 Abdelberi Chaabane, Pere Manils, and Mohamed Ali Kaafar.
Digging into anonymous traﬃc: A deep analysis of the Tor
anonymizing network. In International Conference on Network
and System Security (NSS), pages 167–174. IEEE, 2010.
 Nicholas Hopper. Challenges in protecting Tor hidden services
from botnet abuse. In International Conference on Finan-
cial Cryptography and Data Security, pages 316–325. Springer,
 J¨org Lenhard, Karsten Loesing, and Guido Wirtz. Perfor-
mance measurements of Tor hidden services in low-bandwidth
access networks. In Applied Cryptography and Network Secu-
rity, pages 324–341. Springer, 2009.
 Ian Goldberg. On the security of the Tor authentication pro-
tocol. In Privacy Enhancing Technologies, pages 316–331.
 Rob Jansen, Nicholas Hopper, and Yongdae Kim. Recruiting
new tor relays with braids. In Proceedings of the 17th ACM
conference on Computer and communications security, pages
319–328. ACM, 2010.
 Roger Dingledine, Dan S Wallach, et al. Building incentives
into Tor. In Financial Cryptography and Data Security, pages
238–256. Springer, 2010.
 Qiyan Wang, Zi Lin, Nikita Borisov, and Nicholas Hopper.
rbridge: User Reputation based Tor Bridge Distribution with
Privacy Preservation. In Network and Distributed System Se-
curity Symposium (NDSS), 2013.
 Rob Smits, Divam Jain, Sarah Pidcock, Ian Goldberg, and
Urs Hengartner. Bridgespa: Improving Tor bridges with sin-
gle packet authorization. In Proceedings of the workshop on
Privacy in the electronic society, pages 93–102. ACM, 2011.
 Hooman Mohajeri Moghaddam, Baiyu Li, Mohammad Der-
akhshani, and Ian Goldberg. Skypemorph: Protocol obfus-
cation for Tor bridges. In Proceedings of the conference on
Computer and Communications Security (CCS), pages 97–
108. ACM, 2012.
 Zachary Weinberg, Jeﬀrey Wang, Vinod Yegneswaran, Linda
Briesemeister, Steven Cheung, Frank Wang, and Dan Boneh.
StegoTorus: a camouﬂage proxy for the Tor anonymity system.
In Proceedings of the conference on Computer and Communi-
cations Security (CCS), pages 109–120. ACM, 2012.
 Deepika Gopal and Nadia Heninger. Torchestra: Reducing
interactive traﬃc delays over Tor. In Proceedings of the work-
shop on Privacy in the electronic society, pages 31–42. ACM,
 Mashael AlSabah, Kevin Bauer, Ian Goldberg, Dirk Grunwald,
Damon McCoy, Stefan Savage, and Geoﬀrey M Voelker. Defen-
estraTor: Throwing out windows in Tor. In Privacy Enhancing
Technologies, pages 134–154. Springer, 2011.
 Rob Jansen, Paul F Syverson, and Nicholas Hopper. Throt-
tling Tor Bandwidth Parasites. In USENIX Security Sympo-
sium, pages 349–363, 2012.
 Akshaya Mani and Micah Sherr. Historε: Diﬀerentially pri-
vate and robust statistics collection for tor. In Network and
Distributed System Security Symposium (NDSS), 2017.
 Stephen Doswell, Nauman Aslam, David Kendall, and Graham
Sexton. The novel use of Bridge Relays to provide persistent
Tor connections for mobile devices. In International Sympo-
sium on Personal Indoor and Mobile Radio Communications
(PIMRC), pages 3371–3375. IEEE, 2013.
 Christer Andersson and Andriy Panchenko. Practical anony-
mous communication on the mobile internet using Tor. In
International Conference on Security and Privacy in Commu-
nications Networks and the Workshops (SecureComm), pages
39–48. IEEE, 2007.