Tor is a computer system that enables anonymous communication. The...
Researchers always guessed that government intelligence agencies li...
The pre-alpha version of Tor was launched in September of 2002 ([he...
Mixing, padding and traffic shaping are techniques that, in this co...
The notion of **symmetric-key** comes from **Symmetric-key encrypti...
*Perfect forward secrecy* essentially means that future compromises...
Tor’s anonymity goals are two-folded: - Tor wants to provide anony...
Tor needs to have a substantial set of users in order for it to hav...
Tor is not a completely decentralized peer-to-peer system like Bitc...
End-to-end traffic timing attacks typically involve an entity with ...
This is a very important point to take into consideration when thin...
Diffie–Hellman handshake (D–H) is a method of securely exchanging c...
This used to be a fairly common mistake for Tor users early on. Thi...
When you run a Tor exit node, that computer will be executing reque...
It is hard to know exactly what type of tactics powerful attackers ...
Tor: The Second-Generation Onion Router
Roger Dingledine
The Free Haven Project
arma@freehaven.net
Nick Mathewson
The Free Haven Project
nickm@freehaven.net
Paul Syverson
Naval Research Lab
syverson@itd.nrl.navy.mil
Abstract
We present Tor, a circuit-based low-latency anonymous com-
munication service. This second-generation Onion Routing
perfect forward secrecy, congestion control, directory servers,
integrity checking, conﬁgurable exit policies, and a practi-
cal design for location-hidden services via rendezvous points.
Tor works on the real-world Internet, requires no special priv-
ileges or kernel modiﬁcations, requires little synchronization
or coordination between nodes, and provides a reasonable
tradeoff between anonymity, usability, and efﬁciency. We
brieﬂy describe our experiences with an international network
of more than 30 nodes. We close with a list of open problems
in anonymous communication.
1 Overview
Onion Routing is a distributed overlay network designed to
anonymize TCP-based applications like web browsing, se-
cure shell, and instant messaging. Clients choose a path
through the network and build a circuit, in which each node
(or “onion router” or “OR”) in the path knows its predecessor
and successor, but no other nodes in the circuit. Trafﬁc ﬂows
down the circuit in ﬁxed-size cells, which are unwrapped by a
symmetric key at each node (like the layers of an onion) and
relayed downstream. The Onion Routing project published
several design and analysis papers [27, 41, 48, 49]. While a
wide area Onion Routing network was deployed brieﬂy, the
only long-running public implementation was a fragile proof-
of-concept that ran on a single machine. Even this simple
deployment processed connections from over sixty thousand
distinct IP addresses from all over the world at a rate of about
ﬁfty thousand per day. But many critical design and deploy-
ment issues were never resolved, and the design has not been
updated in years. Here we describe Tor, a protocol for asyn-
chronous, loosely federated onion routers that provides the
following improvements over the old Onion Routing design:
Perfect forward secrecy: In the original Onion Routing
design, a single hostile node could record trafﬁc and later
compromise successive nodes in the circuit and force them
to decrypt it. Rather than using a single multiply encrypted
data structure (an onion) to lay each circuit, Tor now uses an
incremental or telescoping path-building design, where the
initiator negotiates session keys with each successive hop in
the circuit. Once these keys are deleted, subsequently com-
promised nodes cannot decrypt old trafﬁc. As a side beneﬁt,
onion replay detection is no longer necessary, and the process
of building circuits is more reliable, since the initiator knows
when a hop fails and can then try extending to a new node.
Separation of “protocol cleaning” from anonymity:
Onion Routing originally required a separate “application
proxy” for each supported application protocol—most of
which were never written, so many applications were never
supported. Tor uses the standard and near-ubiquitous
SOCKS [32] proxy interface, allowing us to support most
TCP-based programs without modiﬁcation. Tor now relies on
the ﬁltering features of privacy-enhancing application-level
proxies such as Privoxy [39], without trying to duplicate those
features itself.
No mixing, padding, or trafﬁc shaping (yet): Onion
Routing originally called for batching and reordering cells
as they arrived, assumed padding between ORs, and in later
cost were discussed, and trafﬁc shaping algorithms were
theorized [49] to provide good security without expensive
cent research [1] and deployment experience [4] suggest that
this level of resource use is not practical or economical; and
have a proven and convenient design for trafﬁc shaping or
low-latency mixing that improves anonymity against a realis-
tic adversary, we leave these strategies out.
Many TCP streams can share one circuit: Onion Rout-
ing originally built a separate circuit for each application-
level request, but this required multiple public key operations
for every request, and also presented a threat to anonymity
from building so many circuits; see Section 9. Tor multi-
plexes multiple TCP streams along each circuit to improve
efﬁciency and anonymity.
Leaky-pipe circuit topology: Through in-band signaling
within the circuit, Tor initiators can direct trafﬁc to nodes
partway down the circuit. This novel approach allows traf-
ﬁc to exit the circuit from the middle—possibly frustrating
trafﬁc shape and volume attacks based on observing the end
of the circuit. (It also allows for long-range padding if future
research shows this to be worthwhile.)
Congestion control: Earlier anonymity designs do not ad-
dress trafﬁc bottlenecks. Unfortunately, typical approaches to
load balancing and ﬂow control in overlay networks involve
inter-node control communication and global views of trafﬁc.
Tor’s decentralized congestion control uses end-to-end acks
to maintain anonymity while allowing nodes at the edges of
the network to detect congestion or ﬂooding and send less
data until the congestion subsides.
Directory servers: The earlier Onion Routing design
planned to ﬂood state information through the network—an
approach that can be unreliable and complex. Tor takes a
simpliﬁed view toward distributing this information. Cer-
tain more trusted nodes act as directory servers: they provide
signed directories describing known routers and their current
Variable exit policies: Tor provides a consistent mecha-
nism for each node to advertise a policy describing the hosts
and ports to which it will connect. These exit policies are crit-
ical in a volunteer-based distributed infrastructure, because
each operator is comfortable with allowing different types of
trafﬁc to exit from his node.
End-to-end integrity checking: The original Onion Rout-
ing design did no integrity checking on data. Any node on the
circuit could change the contents of data cells as they passed
by—for example, to alter a connection request so it would
connect to a different webserver, or to ‘tag’ encrypted trafﬁc
and look for corresponding corrupted trafﬁc at the network
edges [15]. Tor hampers these attacks by verifying data in-
tegrity before it leaves the network.
Rendezvous points and hidden services: Tor provides an
integrated mechanism for responder anonymity via location-
protected servers. Previous Onion Routing designs included
long-lived “reply onions” that could be used to build circuits
to a hidden server, but these reply onions did not provide for-
ward security, and became useless if any node in the path
went down or rotated its keys. In Tor, clients negotiate ren-
dezvous points to connect with hidden servers; reply onions
are no longer required.
Unlike Freedom [8], Tor does not require OS kernel
patches or network stack support. This prevents us from
anonymizing non-TCP protocols, but has greatly helped our
portability and deployability.
We have implemented all of the above features, including
rendezvous points. Our source code is available under a free
license, and Tor is not covered by the patent that affected dis-
tribution and use of earlier versions of Onion Routing. We
have deployed a wide-area alpha network to test the design, to
get more experience with usability and users, and to provide
a research platform for experimentation. As of this writing,
the network stands at 32 nodes spread over two continents.
We review previous work in Section 2, describe our goals
and assumptions in Section 3, and then address the above list
of improvements in Sections 4, 5, and 6. We summarize in
Section 7 how our design stands up to known attacks, and
talk about our early deployment experiences in Section 8. We
conclude with a list of open problems in Section 9 and future
work for the Onion Routing project in Section 10.
2 Related work
Modern anonymity systems date to Chaum’s Mix-Net de-
sign [10]. Chaum proposed hiding the correspondence be-
tween sender and recipient by wrapping messages in layers
of public-key cryptography, and relaying them through a path
composed of “mixes. Each mix in turn decrypts, delays, and
re-orders messages before relaying them onward.
Subsequent relay-based anonymity designs have diverged
in two main directions. Systems like Babel [28], Mix-
master [36], and Mixminion [15] have tried to maximize
anonymity at the cost of introducing comparatively large
and variable latencies. Because of this decision, these high-
latency networks resist strong global adversaries, but intro-
duce too much lag for interactive tasks like web browsing,
Internet chat, or SSH connections.
Tor belongs to the second category: low-latency designs
that try to anonymize interactive network trafﬁc. These sys-
tems handle a variety of bidirectional protocols. They also
provide more convenient mail delivery than the high-latency
anonymous email networks, because the remote mail server
provides explicit and timely delivery conﬁrmation. But be-
cause these designs typically involve many packets that must
be delivered quickly, it is difﬁcult for them to prevent an at-
tacker who can eavesdrop both ends of the communication
from correlating the timing and volume of trafﬁc entering the
anonymity network with trafﬁc leaving it [45]. These proto-
cols are similarly vulnerable to an active adversary who in-
troduces timing patterns into trafﬁc entering the network and
looks for correlated patterns among exiting trafﬁc. Although
some work has been done to frustrate these attacks, most de-
signs protect primarily against trafﬁc analysis rather than traf-
ﬁc conﬁrmation (see Section 3.1).
The simplest low-latency designs are single-hop proxies
such as the Anonymizer [3]: a single trusted server strips
the data’s origin before relaying it. These designs are easy to
analyze, but users must trust the anonymizing proxy. Concen-
trating the trafﬁc to this single point increases the anonymity
set (the people a given user is hiding among), but it is vul-
nerable if the adversary can observe all trafﬁc entering and
leaving the proxy.
More complex are distributed-trust, circuit-based
anonymizing systems. In these designs, a user estab-
lishes one or more medium-term bidirectional end-to-end
circuits, and tunnels data in ﬁxed-size cells. Establishing
circuits is computationally expensive and typically requires
public-key cryptography, whereas relaying cells is compar-
atively inexpensive and typically requires only symmetric
encryption. Because a circuit crosses several servers, and
each server only knows the adjacent servers in the circuit, no
single server can link a user to her communication partners.
The Java Anon Proxy (also known as JAP or Web MIXes)
uses ﬁxed shared routes known as cascades. As with a
single-hop proxy, this approach aggregates users into larger
anonymity sets, but again an attacker only needs to observe
both ends of the cascade to bridge all the system’s trafﬁc. The
Java Anon Proxy’s design calls for padding between end users
and the head of the cascade [7]. However, it is not demon-
strated whether the current implementation’s padding policy
improves anonymity.
PipeNet [5, 12], another low-latency design proposed
around the same time as Onion Routing, gave stronger
anonymity but allowed a single user to shut down the net-
work by not sending. Systems like ISDN mixes [38] were
designed for other environments with different assumptions.
In P2P designs like Tarzan [24] and MorphMix [43], all
participants both generate trafﬁc and relay trafﬁc for others.
These systems aim to conceal whether a given peer originated
a request or just relayed it from another peer. While Tarzan
and MorphMix use layered encryption as above, Crowds [42]
simply assumes an adversary who cannot observe the initia-
tor: it uses no public-key encryption, so any node on a circuit
Hordes [34] is based on Crowds but also uses multicast
responses to hide the initiator. Herbivore [25] and P
5
[46]
go even further, requiring broadcast. These systems are de-
signed primarily for communication among peers, although
Herbivore users can make external connections by requesting
a peer to serve as a proxy.
Systems like Freedom and the original Onion Routing
build circuits all at once, using a layered “onion” of public-
key encrypted messages, each layer of which provides ses-
sion keys and the address of the next server in the circuit.
Tor as described herein, Tarzan, MorphMix, Cebolla [9],
and Rennhard’s Anonymity Network [44] build circuits in
stages, extending them one hop at a time. Section 4.2 de-
scribes how this approach enables perfect forward secrecy.
Circuit-based designs must choose which protocol layer to
anonymize. They may intercept IP packets directly, and re-
lay them whole (stripping the source address) along the cir-
cuit [8, 24]. Like Tor, they may accept TCP streams and
relay the data in those streams, ignoring the breakdown of
that data into TCP segments [43, 44]. Finally, like Crowds,
they may accept application-level protocols such as HTTP
and relay the application requests themselves. Making this
protocol-layer decision requires a compromise between ﬂexi-
bility and anonymity. For example, a system that understands
HTTP can strip identifying information from requests, can
take advantage of caching to limit the number of requests that
leave the network, and can batch or encode requests to min-
imize the number of connections. On the other hand, an IP-
level anonymizer can handle nearly any protocol, even ones
unforeseen by its designers (though these systems require
kernel-level modiﬁcations to some operating systems, and so
are more complex and less portable). TCP-level anonymity
networks like Tor present a middle approach: they are ap-
plication neutral (so long as the application supports, or can
be tunneled across, TCP), but by treating application connec-
tions as data streams rather than raw TCP packets, they avoid
the inefﬁciencies of tunneling TCP over TCP.
Distributed-trust anonymizing systems need to prevent at-
tackers from adding too many servers and thus compromising
user paths. Tor relies on a small set of well-known directory
servers, run by independent parties, to decide which nodes
can join. Tarzan and MorphMix allow unknown users to run
servers, and use a limited resource (like IP addresses) to pre-
vent an attacker from controlling too much of the network.
Crowds suggests requiring written, notarized requests from
potential crowd members.
Anonymous communication is essential for censorship-
resistant systems like Eternity [2], Free Haven [19], Pub-
lius [53], and Tangler [52]. Tor’s rendezvous points enable
connections between mutually anonymous entities; they are a
building block for location-hidden servers, which are needed
by Eternity and Free Haven.
3 Design goals and assumptions
Goals
Like other low-latency anonymity designs, Tor seeks to frus-
trate attackers from linking communication partners, or from
linking multiple communications to or from a single user.
Within this main goal, however, several considerations have
directed Tor’s evolution.
Deployability: The design must be deployed and used in
the real world. Thus it must not be expensive to run (for
example, by requiring more bandwidth than volunteers are
willing to provide); must not place a heavy liability burden
on operators (for example, by allowing attackers to implicate
onion routers in illegal activities); and must not be difﬁcult
or expensive to implement (for example, by requiring kernel
patches, or separate proxies for every protocol). We also can-
not require non-anonymous parties (such as websites) to run
our software. (Our rendezvous point design does not meet
this goal for non-anonymous users talking to hidden servers,
however; see Section 5.)
Usability: A hard-to-use system has fewer users—and be-
cause anonymity systems hide users among users, a system
with fewer users provides less anonymity. Usability is thus
not only a convenience: it is a security requirement [1, 5].
Tor should therefore not require modifying familiar applica-
tions; should not introduce prohibitive delays; and should re-
quire as few conﬁguration decisions as possible. Finally, Tor
should be easily implementable on all common platforms; we
cannot require users to change their operating system to be
anonymous. (Tor currently runs on Win32, Linux, Solaris,
BSD-style Unix, MacOS X, and probably others.)
Flexibility: The protocol must be ﬂexible and well-
speciﬁed, so Tor can serve as a test-bed for future research.
Many of the open problems in low-latency anonymity net-
works, such as generating dummy trafﬁc or preventing Sybil
attacks [22], may be solvable independently from the issues
solved by Tor. Hopefully future systems will not need to rein-
vent Tor’s design.
Simple design: The protocol’s design and security param-
eters must be well-understood. Additional features impose
implementation and complexity costs; adding unproven
techniques to the design threatens deployability, readability,
and ease of security analysis. Tor aims to deploy a simple and
stable system that integrates the best accepted approaches to
protecting anonymity.
Non-goals
In favoring simple, deployable designs, we have explicitly de-
ferred several possible goals, either because they are solved
elsewhere, or because they are not yet solved.
Not peer-to-peer: Tarzan and MorphMix aim to scale
to completely decentralized peer-to-peer environments with
thousands of short-lived servers, many of which may be con-
trolled by an adversary. This approach is appealing, but still
has many open problems [24, 43].
Not secure against end-to-end attacks: Tor does not
claim to completely solve end-to-end timing or intersection
attacks. Some approaches, such as having users run their own
onion routers, may help; see Section 9 for more discussion.
No protocol normalization: Tor does not provide proto-
col normalization like Privoxy or the Anonymizer. If senders
want anonymity from responders while using complex and
variable protocols like HTTP, Tor must be layered with a
ﬁltering proxy such as Privoxy to hide differences between
clients, and expunge protocol features that leak identity. Note
that by this separation Tor can also provide services that are
anonymous to the network yet authenticated to the responder,
like SSH. Similarly, Tor does not integrate tunneling for non-
stream-based protocols like UDP; this must be provided by
an external service if appropriate.
Not steganographic: Tor does not try to conceal who is
connected to the network.
3.1 Threat Model
A global passive adversary is the most commonly assumed
threat when analyzing theoretical anonymity designs. But
like all practical low-latency systems, Tor does not protect
sary who can observe some fraction of network trafﬁc; who
can generate, modify, delete, or delay trafﬁc; who can oper-
ate onion routers of his own; and who can compromise some
fraction of the onion routers.
In low-latency anonymity systems that use layered encryp-
tion, the adversary’s typical goal is to observe both the ini-
tiator and the responder. By observing both ends, passive at-
tackers can conﬁrm a suspicion that Alice is talking to Bob if
the timing and volume patterns of the trafﬁc on the connec-
tion are distinct enough; active attackers can induce timing
signatures on the trafﬁc to force distinct patterns. Rather than
focusing on these trafﬁc conﬁrmation attacks, we aim to pre-
vent trafﬁc analysis attacks, where the adversary uses trafﬁc
patterns to learn which points in the network he should attack.
communication partners, or try to build a proﬁle of Alice’s
behavior. He might mount passive attacks by observing the
network edges and correlating trafﬁc entering and leaving the
network—by relationships in packet timing, volume, or ex-
ternally visible user-selected options. The adversary can also
mount active attacks by compromising routers or keys; by re-
playing trafﬁc; by selectively denying service to trustworthy
routers to move users to compromised routers, or denying ser-
vice to users to see if trafﬁc elsewhere in the network stops; or
by introducing patterns into trafﬁc that can later be detected.
The adversary might subvert the directory servers to give
users differing views of network state. Additionally, he can
try to decrease the network’s reliability by attacking nodes
or by performing antisocial activities from reliable nodes and
trying to get them taken down—making the network unre-
liable ﬂushes users to other less anonymous systems, where
they may be easier to attack. We summarize in Section 7 how
well the Tor design defends against each of these attacks.
4 The Tor Design
The Tor network is an overlay network; each onion router
(OR) runs as a normal user-level process without any special
privileges. Each onion router maintains a TLS [17] connec-
tion to every other onion router. Each user runs local software
called an onion proxy (OP) to fetch directories, establish cir-
cuits across the network, and handle connections from user
applications. These onion proxies accept TCP streams and
multiplex them across the circuits. The onion router on the
other side of the circuit connects to the requested destinations
and relays data.
Each onion router maintains a long-term identity key and
a short-term onion key. The identity key is used to sign TLS
certiﬁcates, to sign the OR’s router descriptor (a summary of
its keys, address, bandwidth, exit policy, and so on), and (by
directory servers) to sign directories. The onion key is used
to decrypt requests from users to set up a circuit and negotiate
ephemeral keys. The TLS protocol also establishes a short-
term link key when communicating between ORs. Short-term
keys are rotated periodically and independently, to limit the
impact of key compromise.
Section 4.1 presents the ﬁxed-size cells that are the unit
of communication in Tor. We describe in Section 4.2 how
circuits are built, extended, truncated, and destroyed. Sec-
tion 4.3 describes how TCP streams are routed through the
network. We address integrity checking in Section 4.4, and
resource limiting in Section 4.5. Finally, Section 4.6 talks
about congestion control and fairness issues.
4.1 Cells
Onion routers communicate with one another, and with users’
OPs, via TLS connections with ephemeral keys. Using TLS
conceals the data on the connection with perfect forward se-
crecy, and prevents an attacker from modifying data on the
wire or impersonating an OR.
Trafﬁc passes along these connections in ﬁxed-size cells.
Each cell is 512 bytes, and consists of a header and a pay-
speciﬁes which circuit the cell refers to (many circuits can
be multiplexed over the single TLS connection), and a com-
mand to describe what to do with the cell’s payload. (Circuit
identiﬁers are connection-speciﬁc: each circuit has a differ-
ent circID on each OP/OR or OR/OR connection it traverses.)
Based on their command, cells are either control cells, which
are always interpreted by the node that receives them, or re-
lay cells, which carry end-to-end stream data. The control
cell commands are: padding (currently used for keepalive,
set up a new circuit); and destroy (to tear down a circuit).
the front of the payload, containing a streamID (stream iden-
tiﬁer: many streams can be multiplexed over a circuit); an
end-to-end checksum for integrity checking; the length of the
relay payload; and a relay command. The entire contents of
decrypted together as the relay cell moves along the circuit,
using the 128-bit AES cipher in counter mode to generate a
cipher stream. The relay commands are: relay data (for data
ﬂowing down the stream), relay begin (to open a stream), re-
lay end (to close a stream cleanly), relay teardown (to close a
broken stream), relay connected (to notify the OP that a relay
begin has succeeded), relay extend and relay extended (to ex-
tend the circuit by a hop, and to acknowledge), relay truncate
and relay truncated (to tear down only part of the circuit, and
to acknowledge), relay sendme (used for congestion control),
and relay drop (used to implement long-range dummies). We
give a visual overview of cell structure plus the details of re-
lay cell structure, and then describe each of these cell types
and commands in more detail below.
CircID
2
Relay StreamID Digest Len DATA
CircID CMD
2 1
DATA
2
CMD
1
509 bytes
1 2 6 498
4.2 Circuits and streams
Onion Routing originally built one circuit for each TCP
stream. Because building a circuit can take several tenths
of a second (due to public-key cryptography and network la-
tency), this design imposed high costs on applications like
web browsing that open many TCP streams.
In Tor, each circuit can be shared by many TCP streams.
To avoid delays, users construct circuits preemptively. To
limit linkability among their streams, users’ OPs build a new
circuit periodically if the previous ones have been used, and
expire old used circuits that no longer have any open streams.
OPs consider rotating to a new circuit once a minute: thus
even heavy users spend negligible time building circuits, but
a limited number of requests can be linked to each other
through a given exit node. Also, because circuits are built in
the background, OPs can recover from failed circuit creation
without harming user experience.
OR 1Alice OR 2
"HTTP GET..."
. . . . . .. . .
(TCP handshake)
website
{X}−−AES encryption
E(x)−−RSA encryption
Legend:
Relay c1{Extend, OR2, E(g^x2)}
Relay c1{{Begin <website>:80}}
Relay c1{Extended, g^y2, H(K2)}
Relay c2{Begin <website>:80}
Relay c1{{Connected}}
Relay c2{Connected}
Relay c1{{Data, "HTTP GET..."}}
Relay c2{Data, "HTTP GET..."}
cN−−a circID
Relay c1{{Data, (response)}}
(response)
Relay c2{Data, (response)}
Created c2, g^y2, H(K2)
Create c2, E(g^x2)
Create c1, E(g^x1)
Created c1, g^y1, H(K1)
Figure 1: Alice builds a two-hop circuit and begins fetching
a web page.
Constructing a circuit
A user’s OP constructs circuits incrementally, negotiating a
symmetric key with each OR on the circuit, one hop at a time.
To begin creating a new circuit, the OP (call her Alice) sends
a create cell to the ﬁrst node in her chosen path (call him
Bob). (She chooses a new circID C
AB
not currently used on
the connection from her to Bob.) The create cell’s payload
contains the ﬁrst half of the Difﬁe-Hellman handshake (g
x
),
encrypted to the onion key of the OR (call him Bob). Bob
responds with a created cell containing g
y
along with a hash
of the negotiated key K = g
xy
.
Once the circuit has been established, Alice and Bob can
send one another relay cells encrypted with the negotiated
key.
1
More detail is given in the next section.
To extend the circuit further, Alice sends a relay extend cell
to Bob, specifying the address of the next OR (call her Carol),
and an encrypted g
x
2
for her. Bob copies the half-handshake
into a create cell, and passes it to Carol to extend the cir-
cuit. (Bob chooses a new circID C
BC
not currently used on
the connection between him and Carol. Alice never needs to
know this circID; only Bob associates C
AB
on his connec-
tion with Alice to C
BC
on his connection with Carol.) When
Carol responds with a created cell, Bob wraps the payload
into a relay extended cell and passes it back to Alice. Now
the circuit is extended to Carol, and Alice and Carol share a
common key K
2
= g
x
2
y
2
.
To extend the circuit to a third node or beyond, Alice pro-
ceeds as above, always telling the last node in the circuit to
extend one hop further.
This circuit-level handshake protocol achieves unilateral
entity authentication (Alice knows she’s handshaking with
the OR, but the OR doesn’t care who is opening the circuit—
Alice uses no public key and remains anonymous) and unilat-
eral key authentication (Alice and the OR agree on a key, and
Alice knows only the OR learns it). It also achieves forward
secrecy and key freshness. More formally, the protocol is as
follows (where E
P K
Bob
(·) is encryption with Bob’s public
key, H is a secure hash function, and | is concatenation):
Alice Bob : E
P K
Bob
(g
x
)
Bob Alice : g
y
, H(K|“handshake”)
In the second step, Bob proves that it was he who received
g
x
, and who chose y. We use PK encryption in the ﬁrst step
(rather than, say, using the ﬁrst two steps of STS, which has
a signature in the second step) because a single cell is too
small to hold both a public key and a signature. Preliminary
analysis with the NRL protocol analyzer [35] shows this
protocol to be secure (including perfect forward secrecy)
Relay cells
Once Alice has established the circuit (so she shares keys with
each OR on the circuit), she can send relay cells. Upon re-
ceiving a relay cell, an OR looks up the corresponding circuit,
key for that circuit. If the cell is headed away from Alice the
OR then checks whether the decrypted cell has a valid digest
(as an optimization, the ﬁrst two bytes of the integrity check
are zero, so in most cases we can avoid computing the hash).
If valid, it accepts the relay cell and processes it as described
below. Otherwise, the OR looks up the circID and OR for the
next step in the circuit, replaces the circID as appropriate, and
sends the decrypted relay cell to the next OR. (If the OR at
the end of the circuit receives an unrecognized relay cell, an
error has occurred, and the circuit is torn down.)
1
Actually, the negotiated key is used to derive two symmetric keys: one
for each direction.
OPs treat incoming relay cells similarly: they iteratively
shared with each OR on the circuit, from the closest to far-
thest. If at any stage the digest is valid, the cell must have
originated at the OR whose encryption has just been removed.
To construct a relay cell addressed to a given OR, Alice as-
signs the digest, and then iteratively encrypts the cell payload
of each hop up to that OR. Because the digest is encrypted to
a different value at each step, only at the targeted OR will
it have a meaningful value.
2
This leaky pipe circuit topol-
ogy allows Alice’s streams to exit at different ORs on a sin-
gle circuit. Alice may choose different exit points because of
their exit policies, or to keep the ORs from knowing that two
streams originate from the same person.
When an OR later replies to Alice with a relay cell, it en-
it shares with Alice, and sends the cell back toward Alice
along the circuit. Subsequent ORs add further layers of en-
cryption as they relay the cell back to Alice.
To tear down a circuit, Alice sends a destroy control cell.
Each OR in the circuit receives the destroy cell, closes all
streams on that circuit, and passes a new destroy cell forward.
But just as circuits are built incrementally, they can also be
torn down incrementally: Alice can send a relay truncate cell
to a single OR on a circuit. That OR then sends a destroy cell
forward, and acknowledges with a relay truncated cell. Alice
can then extend the circuit to different nodes, without signal-
ing to the intermediate nodes (or a limited observer) that she
has changed her circuit. Similarly, if a node on the circuit
goes down, the adjacent node can send a relay truncated cell
back to Alice. Thus the “break a node and see which circuits
go down” attack [4] is weakened.
4.3 Opening and closing streams
When Alice’s application wants a TCP connection to a given
address and port, it asks the OP (via SOCKS) to make the
connection. The OP chooses the newest open circuit (or cre-
ates one if needed), and chooses a suitable OR on that circuit
to be the exit node (usually the last node, but maybe others
due to exit policy conﬂicts; see Section 6.2.) The OP then
opens the stream by sending a relay begin cell to the exit node,
using a new random streamID. Once the exit node connects
to the remote host, it responds with a relay connected cell.
Upon receipt, the OP sends a SOCKS reply to notify the ap-
plication of its success. The OP now accepts data from the
application’s TCP stream, packaging it into relay data cells
and sending those cells along the circuit to the chosen OR.
There’s a catch to using SOCKS, however—some applica-
tions pass the alphanumeric hostname to the Tor client, while
others resolve it into an IP address ﬁrst and then pass the IP
2
With 48 bits of digest per cell, the probability of an accidental collision
is far lower than the chance of hardware failure.
address to the Tor client. If the application does DNS resolu-
tion ﬁrst, Alice thereby reveals her destination to the remote
DNS server, rather than sending the hostname through the Tor
network to be resolved at the far end. Common applications
like Mozilla and SSH have this ﬂaw.
With Mozilla, the ﬂaw is easy to address: the ﬁltering
HTTP proxy called Privoxy gives a hostname to the Tor
client, so Alice’s computer never does DNS resolution. But
a portable general solution, such as is needed for SSH, is an
open problem. Modifying or replacing the local nameserver
can be invasive, brittle, and unportable. Forcing the resolver
library to prefer TCP rather than UDP is hard, and also has
portability problems. Dynamically intercepting system calls
to the resolver library seems a promising direction. We could
also provide a tool similar to dig to perform a private lookup
through the Tor network. Currently, we encourage the use of
privacy-aware proxies like Privoxy wherever possible.
Closing a Tor stream is analogous to closing a TCP stream:
it uses a two-step handshake for normal operation, or a one-
step handshake for errors. If the stream closes abnormally,
the adjacent node simply sends a relay teardown cell. If the
stream closes normally, the node sends a relay end cell down
the circuit, and the other side responds with its own relay end
cell. Because all relay cells use layered encryption, only the
destination OR knows that a given relay cell is a request to
close a stream. This two-step handshake allows Tor to support
TCP-based applications that use half-closed connections.
4.4 Integrity checking on streams
Because the old Onion Routing design used a stream cipher
without integrity checking, trafﬁc was vulnerable to a mal-
leability attack: though the attacker could not decrypt cells,
any changes to encrypted data would create corresponding
changes to the data leaving the network. This weakness al-
lowed an adversary who could guess the encrypted content to
change a padding cell to a destroy cell; change the destination
change an FTP command from dir to rm *. (Even an ex-
similarly used a stream cipher.)
cannot modify data. Addressing the insider malleability at-
tack, however, is more complex.
We could do integrity checking of the relay cells at each
hop, either by including hashes or by using an authenticating
cipher mode like EAX [6], but there are some problems. First,
these approaches impose a message-expansion overhead at
each hop, and so we would have to either leak the path length
or waste bytes by padding to a maximum path length. Sec-
ond, these solutions can only verify trafﬁc coming from Al-
ice: ORs would not be able to produce suitable hashes for
the intermediate hops, since the ORs on a circuit do not know
the other ORs’ session keys. Third, we have already accepted
that our design is vulnerable to end-to-end timing attacks; so
tagging attacks performed within the circuit provide no addi-
tional information to the attacker.
Thus, we check integrity only at the edges of each stream.
(Remember that in our leaky-pipe circuit topology, a stream’s
edge could be any hop in the circuit.) When Alice negotiates
a key with a new hop, they each initialize a SHA-1 digest with
a derivative of that key, thus beginning with randomness that
only the two of them know. Then they each incrementally
add to the SHA-1 digest the contents of all relay cells they
create, and include with each relay cell the ﬁrst four bytes of
the current digest. Each also keeps a SHA-1 digest of data
To be sure of removing or modifying a cell, the attacker
must be able to deduce the current digest state (which de-
pends on all trafﬁc between Alice and Bob, starting with their
negotiated key). Attacks on SHA-1 where the adversary can
incrementally add to a hash to produce a new valid hash don’t
work, because all hashes are end-to-end encrypted across the
circuit. The computational overhead of computing the digests
is minimal compared to doing the AES encryption performed
at each hop of the circuit. We use only four bytes per cell
rectly guess a valid hash is acceptably low, given that the OP
or OR tear down the circuit if they receive a bad hash.
4.5 Rate limiting and fairness
Volunteers are more willing to run services that can limit
their bandwidth usage. To accommodate them, Tor servers
use a token bucket approach [50] to enforce a long-term aver-
age rate of incoming bytes, while still permitting short-term
bursts above the allowed bandwidth.
Because the Tor protocol outputs about the same number
of bytes as it takes in, it is sufﬁcient in practice to limit only
incoming bytes. With TCP streams, however, the correspon-
dence is not one-to-one: relaying a single incoming byte can
require an entire 512-byte cell. (We can’t just wait for more
bytes, because the local application may be awaiting a reply.)
Therefore, we treat this case as if the entire cell size had been
read, regardless of the cell’s fullness.
cuit’s edges can heuristically distinguish interactive streams
from bulk streams by comparing the frequency with which
they supply cells. We can provide good latency for interactive
streams by giving them preferential service, while still giving
good overall throughput to the bulk streams. Such prefer-
ential treatment presents a possible end-to-end attack, but an
this information through timing attacks.
4.6 Congestion control
Even with bandwidth rate limiting, we still need to worry
about congestion, either accidental or intentional. If enough
users choose the same OR-to-OR connection for their cir-
cuits, that connection can become saturated. For example,
an attacker could send a large ﬁle through the Tor network
to a webserver he runs, and then refuse to read any of the
bytes at the webserver end of the circuit. Without some con-
gestion control mechanism, these bottlenecks can propagate
back through the entire network. We don’t need to reimple-
ment full TCP windows (with sequence numbers, the abil-
ity to drop cells when we’re full and retransmit later, and
so on), because TCP already guarantees in-order delivery of
each cell. We describe our response below.
Circuit-level throttling: To control a circuit’s bandwidth
usage, each OR keeps track of two windows. The packaging
window tracks how many relay data cells the OR is allowed to
package (from incoming TCP streams) for transmission back
to the OP, and the delivery window tracks how many relay
data cells it is willing to deliver to TCP streams outside the
network. Each window is initialized (say, to 1000 data cells).
When a data cell is packaged or delivered, the appropriate
window is decremented. When an OR has received enough
data cells (currently 100), it sends a relay sendme cell towards
the OP, with streamID zero. When an OR receives a relay
sendme cell with streamID zero, it increments its packaging
window. Either of these cells increments the corresponding
window by 100. If the packaging window reaches 0, the OR
stops reading from TCP connections for all streams on the
corresponding circuit, and sends no more relay data cells until
receiving a relay sendme cell.
The OP behaves identically, except that it must track a
packaging window and a delivery window for every OR in
the circuit. If a packaging window reaches 0, it stops reading
from streams destined for that OR.
Stream-level throttling: The stream-level congestion con-
trol mechanism is similar to the circuit-level mechanism. ORs
and OPs use relay sendme cells to implement end-to-end ﬂow
control for individual streams across circuits. Each stream
begins with a packaging window (currently 500 cells), and
increments the window by a ﬁxed value (50) upon receiv-
ing a relay sendme cell. Rather than always returning a relay
sendme cell as soon as enough cells have arrived, the stream-
level congestion control also has to check whether data has
been successfully ﬂushed onto the TCP stream; it sends the
relay sendme cell only when the number of bytes pending to
be ﬂushed is under some threshold (currently 10 cells’ worth).
These arbitrarily chosen parameters seem to give tolerable
throughput and delay; see Section 8.
5 Rendezvous Points and hidden services
Rendezvous points are a building block for location-hidden
services (also known as responder anonymity) in the Tor net-
work. Location-hidden services allow Bob to offer a TCP ser-
vice, such as a webserver, without revealing his IP address.
This type of anonymity protects against distributed DoS at-
tacks: attackers are forced to attack the onion routing network
because they do not know Bob’s IP address.
Our design for location-hidden servers has the following
goals. Access-control: Bob needs a way to ﬁlter incoming
requests, so an attacker cannot ﬂood Bob simply by mak-
ing many connections to him. Robustness: Bob should be
able to maintain a long-term pseudonymous identity even in
the presence of router failure. Bob’s service must not be tied
to a single OR, and Bob must be able to migrate his service
across ORs. Smear-resistance: A social attacker should not
be able to “frame” a rendezvous router by offering an ille-
gal or disreputable location-hidden service and making ob-
servers believe the router created that service. Application-
transparency: Although we require users to run special soft-
ware to access location-hidden servers, we must not require
them to modify their applications.
We provide location-hiding for Bob by allowing him to
advertise several onion routers (his introduction points) as
contact points. He may do this on any robust efﬁcient key-
value lookup system with authenticated updates, such as a
distributed hash table (DHT) like CFS [11].
3
Alice, the client,
chooses an OR as her rendezvous point. She connects to one
of Bob’s introduction points, informs him of her rendezvous
point, and then waits for him to connect to the rendezvous
point. This extra level of indirection helps Bob’s introduc-
tion points avoid problems associated with serving unpopular
ﬁles directly (for example, if Bob serves material that the in-
troduction point’s community ﬁnds objectionable, or if Bob’s
service tends to get attacked by network vandals). The ex-
tra level of indirection also allows Bob to respond to some
requests and ignore others.
5.1 Rendezvous points in Tor
The following steps are performed on behalf of Alice and Bob
by their local OPs; application integration is described more
fully below.
Bob generates a long-term public key pair to identify his
service.
Bob chooses some introduction points, and advertises
with his public key. He can add more later.
Bob builds a circuit to each of his introduction points,
and tells them to wait for requests.
3
Rather than rely on an external infrastructure, the Onion Routing net-
work can run the lookup service itself. Our current implementation provides
a simple lookup system on the directory servers.
Alice learns about Bob’s service out of band (perhaps
Bob told her, or she found it on a website). She retrieves
the details of Bob’s service from the lookup service. If
Alice wants to access Bob’s service anonymously, she
must connect to the lookup service via Tor.
Alice chooses an OR as the rendezvous point (RP) for
her connection to Bob’s service. She builds a circuit
to the RP, and gives it a randomly chosen “rendezvous
Alice opens an anonymous stream to one of Bob’s intro-
duction points, and gives it a message (encrypted with
Bob’s public key) telling it about herself, her RP and ren-
dezvous cookie, and the start of a DH handshake. The
introduction point sends the message to Bob.
If Bob wants to talk to Alice, he builds a circuit to Al-
ice’s RP and sends the rendezvous cookie, the second
half of the DH handshake, and a hash of the session key
they now share. By the same argument as in Section 4.2,
Alice knows she shares the key only with Bob.
The RP connects Alice’s circuit to Bob’s. Note that RP
can’t recognize Alice, Bob, or the data they transmit.
Alice sends a relay begin cell along the circuit. It arrives
at Bob’s OP, which connects to Bob’s webserver.
An anonymous stream has been established, and Alice
and Bob communicate as normal.
When establishing an introduction point, Bob provides the
onion router with the public key identifying his service. Bob
signs his messages, so others cannot usurp his introduction
point in the future. He uses the same public key to establish
the other introduction points for his service, and periodically
refreshes his entry in the lookup service.
The message that Alice gives the introduction point in-
cludes a hash of Bob’s public key and an optional initial au-
thorization token (the introduction point can do prescreening,
for example to block replays). Her message to Bob may in-
clude an end-to-end authorization token so Bob can choose
whether to respond. The authorization tokens can be used
to provide selective access: important users can get uninter-
rupted access. During normal situations, Bob’s service might
simply be offered directly from mirrors, while Bob gives
out tokens to high-priority users. If the mirrors are knocked
down, those users can switch to accessing Bob’s service via
the Tor rendezvous system.
Bob’s introduction points are themselves subject to DoS—
he must open many introduction points or risk such an at-
tack. He can provide selected users with a current list or fu-
ture schedule of unadvertised introduction points; this is most
practical if there is a stable and large group of introduction
points available. Bob could also give secret public keys for
consulting the lookup service. All of these approaches limit
exposure even when some selected users collude in the DoS.
5.2 Integration with user applications
Bob conﬁgures his onion proxy to know the local IP address
and port of his service, a strategy for authorizing clients, and
his public key. The onion proxy anonymously publishes a
signed statement of Bob’s public key, an expiration time, and
the current introduction points for his service onto the lookup
service, indexed by the hash of his public key. Bob’s web-
server is unmodiﬁed, and doesn’t even know that it’s hidden
behind the Tor network.
Alice’s applications also work unchanged—her client
interface remains a SOCKS proxy. We encode all of
the necessary information into the fully qualiﬁed domain
name (FQDN) Alice uses when establishing her connection.
Location-hidden services use a virtual top level domain called
.onion: thus hostnames take the form x.y.onion where
x is the authorization cookie and y encodes the hash of
the public key. Alice’s onion proxy examines addresses; if
they’re destined for a hidden server, it decodes the key and
starts the rendezvous as described above.
5.3 Previous rendezvous work
Rendezvous points in low-latency anonymity systems were
ﬁrst described for use in ISDN telephony [30, 38]. Later low-
latency designs used rendezvous points for hiding location
of mobile phones and low-power location trackers [23, 40].
Rendezvous for anonymizing low-latency Internet connec-
tions was suggested in early Onion Routing work [27], but
the ﬁrst published design was by Ian Goldberg [26]. His de-
sign differs from ours in three ways. First, Goldberg suggests
that Alice should manually hunt down a current location of
the service via Gnutella; our approach makes lookup trans-
parent to the user, as well as faster and more robust. Second,
in Tor the client and server negotiate session keys with Difﬁe-
Hellman, so plaintext is not exposed even at the rendezvous
point. Third, our design minimizes the exposure from run-
ning the service, to encourage volunteers to offer introduc-
tion and rendezvous services. Tor’s introduction points do not
output any bytes to the clients; the rendezvous points don’t
know the client or the server, and can’t read the data being
transmitted. The indirection scheme is also designed to in-
clude authentication/authorization—if Alice doesn’t include
the right cookie with her request for service, Bob need not
even acknowledge his existence.
6 Other design decisions
6.1 Denial of service
Providing Tor as a public service creates many opportuni-
ties for denial-of-service attacks against the network. While
ﬂow control and rate limiting (discussed in Section 4.6) pre-
vent users from consuming more bandwidth than routers are
willing to provide, opportunities remain for users to consume
more network resources than their fair share, or to render the
network unusable for others.
First of all, there are several CPU-consuming denial-of-
service attacks wherein an attacker can force an OR to per-
form expensive cryptographic operations. For example, an at-
tacker can fake the start of a TLS handshake, forcing the OR
to carry out its (comparatively expensive) half of the hand-
shake at no real computational cost to the attacker.
We have not yet implemented any defenses for these at-
tacks, but several approaches are possible. First, ORs can
require clients to solve a puzzle [16] while beginning new
TLS handshakes or accepting create cells. So long as these
tokens are easy to verify and computationally expensive to
produce, this approach limits the attack multiplier. Addition-
ally, ORs can limit the rate at which they accept create cells
and TLS connections, so that the computational work of pro-
cessing them does not drown out the symmetric cryptography
operations that keep cells ﬂowing. This rate limiting could,
however, allow an attacker to slow down other users when
they build new circuits.
Adversaries can also attack the Tor network’s hosts and
streams passing along that part of the circuit. Users simi-
larly lose service when a router crashes or its operator restarts
it. The current Tor design treats such attacks as intermit-
tent network failures, and depends on users and applications
to respond or recover as appropriate. A future design could
use an end-to-end TCP-like acknowledgment protocol, so no
streams are lost unless the entry or exit point is disrupted.
This solution would require more buffering at the network
edges, however, and the performance and anonymity impli-
cations from this extra complexity still require investigation.
6.2 Exit policies and abuse
Exit abuse is a serious barrier to wide-scale Tor deployment.
Anonymity presents would-be vandals and abusers with an
opportunity to hide the origins of their activities. Attackers
can harm the Tor network by implicating exit servers for their
abuse. Also, applications that commonly use IP-based au-
thentication (such as institutional mail or webservers) can be
fooled by the fact that anonymous connections appear to orig-
inate at the exit OR.
We stress that Tor does not enable any new class of abuse.
sands of misconﬁgured systems worldwide, and the Tor net-
work is far from the easiest way to launch attacks. But be-
cause the onion routers can be mistaken for the originators
of the abuse, and the volunteers who run them may not want
to deal with the hassle of explaining anonymity networks to
irate administrators, we must block or limit abuse through the
Tor network.
To mitigate abuse issues, each onion router’s exit policy de-
scribes to which external addresses and ports the router will
connect. On one end of the spectrum are open exit nodes
that will connect anywhere. On the other end are middleman
nodes that only relay trafﬁc to other Tor nodes, and private
exit nodes that only connect to a local host or network. A
private exit can allow a client to connect to a given host or
network more securely—an external adversary cannot eaves-
drop trafﬁc between the private exit and the ﬁnal destination,
and so is less sure of Alice’s destination and activities. Most
onion routers in the current network function as restricted ex-
its that permit connections to the world at large, but prevent
SMTP. The OR might also be able to authenticate clients to
prevent exit abuse without harming anonymity [48].
Many administrators use port restrictions to support only a
limited set of services, such as HTTP, SSH, or AIM. This is
not a complete solution, of course, since abuse opportunities
for these protocols are still well known.
We have not yet encountered any abuse in the deployed
network, but if we do we should consider using proxies to
clean trafﬁc for certain protocols as it leaves the network. For
example, much abusive HTTP behavior (such as exploiting
buffer overﬂows or well-known script vulnerabilities) can be
detected in a straightforward manner. Similarly, one could
run automatic spam ﬁltering software (such as SpamAssas-
sin) on email exiting the OR network.
ORs may also rewrite exiting trafﬁc to append headers
or other information indicating that the trafﬁc has passed
through an anonymity service. This approach is commonly
used by email-only anonymity systems. ORs can also run
on servers with hostnames like anonymous to further alert
abuse targets to the nature of the anonymous trafﬁc.
A mixture of open and restricted exit nodes allows the most
ﬂexibility for volunteers running servers. But while having
many middleman nodes provides a large and robust network,
having only a few exit nodes reduces the number of points an
adversary needs to monitor for trafﬁc analysis, and places a
greater burden on the exit nodes. This tension can be seen in
the Java Anon Proxy cascade model, wherein only one node
versary only needs to observe the entry and exit of a cascade
to perform trafﬁc analysis on all that cascade’s users. The hy-
dra model (many entries, few exits) presents a different com-
promise: only a few exit nodes are needed, but an adversary
needs to work harder to watch all the clients; see Section 10.
Finally, we note that exit abuse must not be dismissed as
a peripheral issue: when a system’s public image suffers, it
can reduce the number and diversity of that system’s users,
and thereby reduce the anonymity of the system itself. Like
usability, public perception is a security parameter. Sadly,
preventing abuse of open exit nodes is an unsolved problem,
and will probably remain an arms race for the foreseeable
future. The abuse problems faced by Princeton’s CoDeeN
project [37] give us a glimpse of likely issues.
6.3 Directory Servers
First-generation Onion Routing designs [8, 41] used in-band
network status updates: each router ﬂooded a signed state-
ment to its neighbors, which propagated it onward. But
anonymizing networks have different security goals than typ-
ical link-state routing protocols. For example, delays (acci-
dental or intentional) that can cause different parts of the net-
work to have different views of link-state and topology are
not only inconvenient: they give attackers an opportunity to
exploit differences in client knowledge. We also worry about
attacks to deceive a client about the router membership list,
topology, or current network state. Such partitioning attacks
on client knowledge help an adversary to efﬁciently deploy
resources against a target [15].
Tor uses a small group of redundant, well-known onion
routers to track changes in network topology and node state,
including keys and exit policies. Each such directory server
acts as an HTTP server, so clients can fetch current network
state and router lists, and so other ORs can upload state infor-
mation. Onion routers periodically publish signed statements
of their state to each directory server. The directory servers
combine this information with their own views of network
liveness, and generate a signed description (a directory) of
the entire network state. Client software is pre-loaded with a
list of the directory servers and their keys, to bootstrap each
client’s view of the network.
When a directory server receives a signed statement for an
OR, it checks whether the OR’s identity key is recognized.
Directory servers do not advertise unrecognized ORs—if they
did, an adversary could take over the network by creating
many servers [22]. Instead, new nodes must be approved by
the directory server administrator before they are included.
Mechanisms for automated node approval are an area of ac-
tive research, and are discussed more in Section 9.
Of course, a variety of attacks remain. An adversary who
controls a directory server can track clients by providing them
different information—perhaps by listing only nodes under
its control, or by informing only certain clients about a given
node. Even an external adversary can exploit differences in
client knowledge: clients who use a node listed on one direc-
tory server but not the others are vulnerable.
Thus these directory servers must be synchronized and
redundant, so that they can agree on a common directory.
Clients should only trust this directory if it is signed by a
threshold of the directory servers.
The directory servers in Tor are modeled after those in
Mixminion [15], but our situation is easier. First, we make
the simplifying assumption that all participants agree on the
set of directory servers. Second, while Mixminion needs
to predict node behavior, Tor only needs a threshold con-
sensus of the current state of the network. Third, we as-
sume that we can fall back to the human administrators to
discover and resolve problems when a consensus directory
cannot be reached. Since there are relatively few directory
servers (currently 3, but we expect as many as 9 as the net-
work scales), we can afford operations like broadcast to sim-
plify the consensus-building protocol.
To avoid attacks where a router connects to all the direc-
tory servers but refuses to relay trafﬁc from other routers,
the directory servers must also build circuits and use them to
anonymously test router reliability [18]. Unfortunately, this
defense is not yet designed or implemented.
Using directory servers is simpler and more ﬂexible than
ﬂooding. Flooding is expensive, and complicates the analysis
when we start experimenting with non-clique network topolo-
gies. Signed directories can be cached by other onion routers,
so directory servers are not a performance bottleneck when
we have many users, and do not aid trafﬁc analysis by forcing
clients to announce their existence to any central point.
7 Attacks and Defenses
Below we summarize a variety of attacks, and discuss how
well our design withstands them.
Passive attacks
Observing user trafﬁc patterns. Observing a user’s connec-
tion will not reveal her destination or data, but it will reveal
trafﬁc patterns (both sent and received). Proﬁling via user
connection patterns requires further processing, because mul-
tiple application streams may be operating simultaneously or
in series over a single circuit.
Observing user content. While content at the user end is
encrypted, connections to responders may not be (indeed, the
responding website itself may be hostile). While ﬁltering
content is not a primary goal of Onion Routing, Tor can di-
rectly use Privoxy and related ﬁltering services to anonymize
application data streams.
Option distinguishability. We allow clients to choose con-
ﬁguration options. For example, clients concerned about re-
quest linkability should rotate circuits more often than those
concerned about traceability. Allowing choice may attract
users with different needs; but clients who are in the minor-
ity may lose more anonymity by appearing distinct than they
gain by optimizing their behavior [1].
End-to-end timing correlation. Tor only minimally hides
such correlations. An attacker watching patterns of trafﬁc at
the initiator and the responder will be able to conﬁrm the cor-
respondence with high probability. The greatest protection
currently available against such conﬁrmation is to hide the
connection between the onion proxy and the ﬁrst Tor node,
by running the OP on the Tor node or behind a ﬁrewall. This
approach requires an observer to separate trafﬁc originating at
the onion router from trafﬁc passing through it: a global ob-
server can do this, but it might be beyond a limited observer’s
capabilities.
End-to-end size correlation. Simple packet counting will
also be effective in conﬁrming endpoints of a stream. How-
ever, even without padding, we may have some limited pro-
tection: the leaky pipe topology means different numbers of
packets may enter one end of a circuit than exit at the other.
Website ﬁngerprinting. All the effective passive attacks
above are trafﬁc conﬁrmation attacks, which puts them out-
side our design goals. There is also a passive trafﬁc analysis
attack that is potentially effective. Rather than searching
exit connections for timing and volume correlations, the
adversary may build up a database of “ﬁngerprints” contain-
ing ﬁle sizes and access patterns for targeted websites. He
can later conﬁrm a user’s connection to a given site simply
by consulting the database. This attack has been shown to
be effective against SafeWeb [29]. It may be less effective
against Tor, since streams are multiplexed within the same
circuit, and ﬁngerprinting will be limited to the granularity
of cells (currently 512 bytes). Additional defenses could
include larger cell sizes, padding schemes to group websites
4
Active attacks
Compromise keys. An attacker who learns the TLS session
key can see control cells and encrypted relay cells on every
circuit on that connection; learning a circuit session key lets
him unwrap one layer of the encryption. An attacker who
learns an OR’s TLS private key can impersonate that OR for
the TLS key’s lifetime, but he must also learn the onion key
to decrypt create cells (and because of perfect forward se-
crecy, he cannot hijack already established circuits without
also compromising their session keys). Periodic key rotation
limits the window of opportunity for these attacks. On the
other hand, an attacker who learns a node’s identity key can
replace that node indeﬁnitely by sending new forged descrip-
tors to the directory servers.
Iterated compromise. A roving adversary who can com-
promise ORs (by system intrusion, legal coercion, or extrale-
gal coercion) could march down the circuit compromising the
nodes until he reaches the end. Unless the adversary can com-
plete this attack within the lifetime of the circuit, however,
the ORs will have discarded the necessary information before
the attack can be completed. (Thanks to the perfect forward
secrecy of session keys, the attacker cannot force nodes to de-
crypt recorded trafﬁc once the circuits have been closed.) Ad-
ditionally, building circuits that cross jurisdictions can make
legal coercion harder—this phenomenon is commonly called
“jurisdictional arbitrage. The Java Anon Proxy project re-
cently experienced the need for this approach, when a Ger-
man court forced them to add a backdoor to their nodes [51].
Run a recipient. An adversary running a webserver trivially
4
Note that this ﬁngerprinting attack should not be confused with the much
more complicated latency attacks of [5], which require a ﬁngerprint of the
latencies of all circuits through the network, combined with those from the
network edges to the target user and the responder website.
learns the timing patterns of users connecting to it, and can in-
troduce arbitrary patterns in its responses. End-to-end attacks
become easier: if the adversary can induce users to connect
to his webserver (perhaps by advertising content targeted to
those users), he now holds one end of their connection. There
is also a danger that application protocols and associated pro-
grams can be induced to reveal information about the initiator.
Tor depends on Privoxy and similar protocol cleaners to solve
this latter problem.
Run an onion proxy. It is expected that end users will nearly
always run their own local onion proxy. However, in some
settings, it may be necessary for the proxy to run remotely—
typically, in institutions that want to monitor the activity of
those connecting to the proxy. Compromising an onion proxy
compromises all future connections through it.
DoS non-observed nodes. An observer who can only watch
some of the Tor network can increase the value of this trafﬁc
by attacking non-observed nodes to shut them down, reduce
their reliability, or persuade users that they are not trustwor-
thy. The best defense here is robustness.
Run a hostile OR. In addition to being a local observer, an
isolated hostile node can create circuits through itself, or alter
trafﬁc patterns to affect trafﬁc at other nodes. Nonetheless, a
hostile node must be immediately adjacent to both endpoints
to compromise the anonymity of a circuit. If an adversary can
run multiple ORs, and can persuade the directory servers that
those ORs are trustworthy and independent, then occasionally
some user will choose one of those ORs for the start and an-
other as the end of a circuit. If an adversary controls m > 1
of N nodes, he can correlate at most
m
N
2
of the trafﬁc—
although an adversary could still attract a disproportionately
large amount of trafﬁc by running an OR with a permissive
exit policy, or by degrading the reliability of other routers.
Introduce timing into messages. This is simply a stronger
version of passive timing attacks already discussed earlier.
Tagging attacks. A hostile node could “tag” a cell by al-
tering it. If the stream were, for example, an unencrypted
request to a Web site, the garbled content coming out at the
appropriate time would conﬁrm the association. However, in-
tegrity checks on cells prevent this attack.
Replace contents of unauthenticated protocols. When re-
laying an unauthenticated protocol like HTTP, a hostile exit
node can impersonate the target server. Clients should prefer
protocols with end-to-end authentication.
Replay attacks. Some anonymity protocols are vulnerable
to replay attacks. Tor is not; replaying one side of a hand-
shake will result in a different negotiated session key, and so
the rest of the recorded session can’t be used.
Smear attacks. An attacker could use the Tor network for
socially disapproved acts, to bring the network into disrepute
and get its operators to shut it down. Exit policies reduce
the possibilities for abuse, but ultimately the network requires
volunteers who can tolerate some political heat.
Distribute hostile code. An attacker could trick users
into running subverted Tor software that did not, in fact,
anonymize their connections—or worse, could trick ORs
into running weakened software that provided users with
less anonymity. We address this problem (but do not solve it
completely) by signing all Tor releases with an ofﬁcial public
key, and including an entry in the directory that lists which
versions are currently believed to be secure. To prevent an
attacker from subverting the ofﬁcial release itself (through
threats, bribery, or insider attacks), we provide all releases in
source code form, encourage source audits, and frequently
warn our users never to trust any software (even from us) that
comes without source.
Directory attacks
Destroy directory servers. If a few directory servers disap-
pear, the others still decide on a valid directory. So long
as any directory servers remain in operation, they will still
broadcast their views of the network and generate a consensus
directory. (If more than half are destroyed, this directory will
not, however, have enough signatures for clients to use it au-
tomatically; human intervention will be necessary for clients
to decide whether to trust the resulting directory.)
Subvert a directory server. By taking over a directory
server, an attacker can partially inﬂuence the ﬁnal directory.
Since ORs are included or excluded by majority vote, the cor-
rupt directory can at worst cast a tie-breaking vote to decide
whether to include marginal ORs. It remains to be seen how
often such marginal cases occur in practice.
Subvert a majority of directory servers. An adversary who
controls more than half the directory servers can include as
many compromised ORs in the ﬁnal directory as he wishes.
We must ensure that directory server operators are indepen-
dent and attack-resistant.
Encourage directory server dissent. The directory agree-
ment protocol assumes that directory server operators agree
on the set of directory servers. An adversary who can per-
suade some of the directory server operators to distrust one
another could split the quorum into mutually hostile camps,
thus partitioning users based on which directory they use. Tor
Trick the directory servers into listing a hostile OR. Our
threat model explicitly assumes directory server operators
will be able to ﬁlter out most hostile ORs.
Convince the directories that a malfunctioning OR is
working. In the current Tor implementation, directory servers
assume that an OR is running correctly if they can start a
TLS connection to it. A hostile OR could easily subvert this
test by accepting TLS connections from ORs but ignoring all
cells. Directory servers must actively test ORs by building
circuits and streams as appropriate. The tradeoffs of a similar
approach are discussed in [18].
Attacks against rendezvous points
Make many introduction requests. An attacker could try to
deny Bob service by ﬂooding his introduction points with re-
quests. Because the introduction points can block requests
that lack authorization tokens, however, Bob can restrict the
volume of requests he receives, or require a certain amount of
computation for every request he receives.
Attack an introduction point. An attacker could disrupt a
location-hidden service by disabling its introduction points.
But because a service’s identity is attached to its public key,
the service can simply re-advertise itself at a different intro-
that only high-priority clients know the address of Bob’s in-
troduction points or so that different clients know of different
introduction points. This forces the attacker to disable all pos-
sible introduction points.
Compromise an introduction point. An attacker who con-
trols Bob’s introduction point can ﬂood Bob with introduction
requests, or prevent valid introduction requests from reaching
him. Bob can notice a ﬂood, and close the circuit. To notice
blocking of valid requests, however, he should periodically
test the introduction point by sending rendezvous requests
and making sure he receives them.
Compromise a rendezvous point. A rendezvous point is no
more sensitive than any other OR on a circuit, since all data
passing through the rendezvous is encrypted with a session
key shared by Alice and Bob.
8 Early experiences: Tor in the Wild
As of mid-May 2004, the Tor network consists of 32 nodes
(24 in the US, 8 in Europe), and more are joining each week
as the code matures. (For comparison, the current remailer
network has about 40 nodes.) Each node has at least a
768Kb/768Kb connection, and many have 10Mb. The num-
ber of users varies (and of course, it’s hard to tell for sure), but
we sometimes have several hundred users—administrators at
several companies have begun sending their entire depart-
ments’ web trafﬁc through Tor, to block other divisions of
their company from reading their trafﬁc. Tor users have re-
ported using the network for web browsing, FTP, IRC, AIM,
Kazaa, SSH, and recipient-anonymous email via rendezvous
points. One user has anonymously set up a Wiki as a hidden
service, where other users anonymously publish the addresses
of their hidden services.
Each Tor node currently processes roughly 800,000 relay
cells (a bit under half a gigabyte) per week. On average, about
80% of each 498-byte payload is full for cells going back to
the client, whereas about 40% is full for cells coming from the
client. (The difference arises because most of the network’s
trafﬁc is web browsing.) Interactive trafﬁc like SSH brings
down the average a lot—once we have more experience, and
assuming we can resolve the anonymity issues, we may parti-
tion trafﬁc into two relay cell sizes: one to handle bulk trafﬁc
and one for interactive trafﬁc.
Based in part on our restrictive default exit policy (we re-
ject SMTP requests) and our low proﬁle, we have had no
abuse issues since the network was deployed in October 2003.
Our slow growth rate gives us time to add features, resolve
bugs, and get a feel for what users actually want from an
anonymity system. Even though having more users would
bolster our anonymity sets, we are not eager to attract the
Kazaa or warez communities—we feel that we must build a
reputation for privacy, human rights, research, and other so-
cially laudable activities.
As for performance, proﬁling shows that Tor spends almost
all its CPU time in AES, which is fast. Current latency is
attributable to two factors. First, network latency is critical:
we are intentionally bouncing trafﬁc around the world several
times. Second, our end-to-end congestion control algorithm
focuses on protecting volunteer servers from accidental DoS
rather than on optimizing performance. To quantify these ef-
fects, we did some informal tests using a network of 4 nodes
on the same machine (a heavily loaded 1GHz Athlon). We
minutes for 54 hours (108 sample points). It arrived in about
300 seconds on average, compared to 210s for a direct down-
load. We ran a similar test on the production Tor network,
fetching the front page of cnn.com (55 kilobytes): while
0.4s, with a median at 2.8s, and 90% ﬁnishing within 5.3s. It
seems that as the network expands, the chance of building a
slow circuit (one that includes a slow or heavily loaded node
or link) is increasing. On the other hand, as our users remain
satisﬁed with this increased latency, we can address our per-
formance incrementally as we proceed with development.
Although Tor’s clique topology and full-visibility directo-
ries present scaling problems, we still expect the network to
support a few hundred nodes and maybe 10,000 users before
we’re forced to become more distributed. With luck, the ex-
perience we gain running the current topology will help us
choose among alternatives when the time comes.
9 Open Questions in Low-latency Anonymity
In addition to the non-goals in Section 3, many questions
must be solved before we can be conﬁdent of Tor’s security.
Many of these open issues are questions of balance. For
example, how often should users rotate to fresh circuits? Fre-
quent rotation is inefﬁcient, expensive, and may lead to inter-
section attacks and predecessor attacks [54], but infrequent
rotation makes the user’s trafﬁc linkable. Besides opening
fresh circuits, clients can also exit from the middle of the cir-
cuit, or truncate and re-extend the circuit. More analysis is
needed to determine the proper tradeoff.
How should we choose path lengths? If Alice always uses
two hops, then both ORs can be certain that by colluding they
will learn about Alice and Bob. In our current approach, Alice
always chooses at least three nodes unrelated to herself and
her destination. Should Alice choose a random path length
(e.g. from a geometric distribution) to foil an attacker who
uses timing to learn that he is the ﬁfth hop and thus concludes
that both Alice and the responder are running ORs?
Throughout this paper, we have assumed that end-to-end
trafﬁc conﬁrmation will immediately and automatically de-
feat a low-latency anonymity system. Even high-latency
anonymity systems can be vulnerable to end-to-end trafﬁc
conﬁrmation, if the trafﬁc volumes are high enough, and if
users’ habits are sufﬁciently distinct [14, 31]. Can anything
be done to make low-latency systems resist these attacks as
well as high-latency systems? Tor already makes some ef-
fort to conceal the starts and ends of streams by wrapping
long-range control commands in identical-looking relay cells.
packets; long-range padding could work against observers
who own the ﬁrst hop in a circuit. But more research remains
to ﬁnd an efﬁcient and practical approach. Volunteers pre-
fer not to run constant-bandwidth padding; but no convinc-
ing trafﬁc shaping approach has been speciﬁed. Recent work
on long-range padding [33] shows promise. One could also
try to reduce correlation in packet timing by batching and re-
ordering packets, but it is unclear whether this could improve
anonymity without introducing so much latency as to render
the network unusable.
A cascade topology may better defend against trafﬁc con-
ﬁrmation by aggregating users, and making padding and mix-
ing more affordable. Does the hydra topology (many input
nodes, few output nodes) work better against some adver-
saries? Are we going to get a hydra anyway because most
nodes will be middleman nodes?
Common wisdom suggests that Alice should run her own
OR for best anonymity, because trafﬁc coming from her node
could plausibly have come from elsewhere. How much mix-
ing does this approach need? Is it immediately beneﬁcial
because of real-world adversaries that can’t observe Alice’s
router, but can run routers of their own?
To scale to many users, and to prevent an attacker from
observing the whole network, it may be necessary to support
far more servers than Tor currently anticipates. This intro-
duces several issues. First, if approval by a central set of di-
rectory servers is no longer feasible, what mechanism should
be used to prevent adversaries from signing up many collud-
ing servers? Second, if clients can no longer have a complete
picture of the network, how can they perform discovery while
preventing attackers from manipulating or exploiting gaps in
their knowledge? Third, if there are too many servers for ev-
ery server to constantly communicate with every other, which
non-clique topology should the network use? (Restricted-
route topologies promise comparable anonymity with better
scalability [13], but whatever topology we choose, we need
some way to keep attackers from manipulating their posi-
tion within it [21].) Fourth, if no central authority is track-
ing server reliability, how do we stop unreliable servers from
making the network unusable? Fifth, do clients receive so
much anonymity from running their own ORs that we should
expect them all to do so [1], or do we need another incentive
structure to motivate them? Tarzan and MorphMix present
possible solutions.
When a Tor node goes down, all its circuits (and thus
streams) must break. Will users abandon the system be-
cause of this brittleness? How well does the method in Sec-
tion 6.1 allow streams to survive node failure? If affected
users rebuild circuits immediately, how much anonymity is
lost? It seems the problem is even worse in a peer-to-peer
environment—such systems don’t yet provide an incentive
for peers to stay connected when they’re done retrieving con-
tent, so we would expect a higher churn rate.
10 Future Directions
Tor brings together many innovations into a uniﬁed deploy-
able system. The next immediate steps include:
Scalability: Tor’s emphasis on deployability and design
simplicity has led us to adopt a clique topology, semi-
centralized directories, and a full-network-visibility model
for client knowledge. These properties will not scale past
a few hundred servers. Section 9 describes some promising
approaches, but more deployment experience will be helpful
in learning the relative importance of these bottlenecks.
Bandwidth classes: This paper assumes that all ORs have
MorphMix model, where nodes advertise their bandwidth
level (DSL, T1, T3), and Alice avoids bottlenecks by choos-
ing nodes that match or exceed her bandwidth. In this way
DSL users can usefully join the Tor network.
Incentives: Volunteers who run nodes are rewarded with
publicity and possibly better anonymity [1]. More nodes
means increased scalability, and more users can mean more
anonymity. We need to continue examining the incentive
structures for participating in Tor. Further, we need to ex-
plore more approaches to limiting abuse, and understand why
most people don’t bother using privacy systems.
Cover trafﬁc: Currently Tor omits cover trafﬁc—its costs
in performance and bandwidth are clear but its security ben-
eﬁts are not well understood. We must pursue more research
on link-level cover trafﬁc and long-range cover trafﬁc to de-
termine whether some simple padding method offers provable
Caching at exit nodes: Perhaps each exit node should run
a caching web proxy [47], to improve anonymity for cached
pages (Alice’s request never leaves the Tor network), to im-
prove speed, and to reduce bandwidth cost. On the other
hand, forward security is weakened because caches consti-
tute a record of retrieved ﬁles. We must ﬁnd the right balance
between usability and security.
description of the entire network every 15 minutes. As the
state grows larger and clients more numerous, we may need
rectory state. More generally, we must ﬁnd more scalable yet
practical ways to distribute up-to-date snapshots of network
status without introducing new attacks.
Further speciﬁcation review: Our public byte-level spec-
iﬁcation [20] needs external review. We hope that as Tor is
deployed, more people will examine its speciﬁcation.
Multisystem interoperability: We are currently working
with the designer of MorphMix to unify the speciﬁcation and
implementation of the common elements of our two systems.
So far, this seems to be relatively straightforward. Interop-
erability will allow testing and direct comparison of the two
designs for trust and scalability.
Wider-scale deployment: The original goal of Tor was to
gain experience in deploying an anonymizing overlay net-
work, and learn from having actual users. We are now at a
point in design and development where we can start deploy-
ing a wider network. Once we have many actual users, we
will doubtlessly be better able to evaluate some of our design
decisions, including our robustness/latency tradeoffs, our per-
formance tradeoffs (including cell size), our abuse-prevention
mechanisms, and our overall usability.
Acknowledgments
Joseph Sokol-Margolis, John Bashinski, and Zack Brown for
editing and comments; Matej Pfajfar, Andrei Serjantov, Marc
Rennhard for design discussions; Bram Cohen for congestion
control discussions; Adam Back for suggesting telescoping
circuits; and Cathy Meadows for formal analysis of the ex-
tend protocol. This work has been supported by ONR and
DARPA.
References
[1] A. Acquisti, R. Dingledine, and P. Syverson. On the eco-
nomics of anonymity. In R. N. Wright, editor, Financial Cryp-
tography. Springer-Verlag, LNCS 2742, 2003.
[2] R. Anderson. The eternity service. In Pragocrypt ’96, 1996.
[3] The Anonymizer. <http://anonymizer.com/>.
[4] A. Back, I. Goldberg, and A. Shostack. Freedom systems 2.1
security issues and analysis. White paper, Zero Knowledge
Systems, Inc., May 2001.
[5] A. Back, U. M
¨
oller, and A. Stiglic. Trafﬁc analysis at-
tacks and trade-offs in anonymity providing systems. In I. S.
Moskowitz, editor, Information Hiding (IH 2001), pages 245–
257. Springer-Verlag, LNCS 2137, 2001.
[6] M. Bellare, P. Rogaway, and D. Wagner. The EAX mode of
operation: A two-pass authenticated-encryption scheme opti-
mized for simplicity and efﬁciency. In Fast Software Encryp-
tion 2004, February 2004.
[7] O. Berthold, H. Federrath, and S. K
¨
opsell. Web MIXes: A
system for anonymous and unobservable Internet access. In
H. Federrath, editor, Designing Privacy Enhancing Technolo-
gies: Workshop on Design Issue in Anonymity and Unobserv-
ability. Springer-Verlag, LNCS 2009, 2000.
[8] P. Boucher, A. Shostack, and I. Goldberg. Freedom systems
2.0 architecture. White paper, Zero Knowledge Systems, Inc.,
December 2000.
[9] Z. Brown. Cebolla: Pragmatic IP Anonymity. In Ottawa Linux
Symposium, June 2002.
[10] D. Chaum. Untraceable electronic mail, return addresses,
and digital pseudo-nyms. Communications of the ACM, 4(2),
February 1981.
[11] F. Dabek, M. F. Kaashoek, D. Karger, R. Morris, and I. Stoica.
Wide-area cooperative storage with CFS. In 18th ACM Sym-
posium on Operating Systems Principles (SOSP ’01), Chateau
Lake Louise, Banff, Canada, October 2001.
[12] W. Dai. Pipenet 1.1. Usenet post, August 1996. <http:
//www.eskimo.com/˜weidai/pipenet.txt> First
mentioned in a post to the cypherpunks list, Feb. 1995.
[13] G. Danezis. Mix-networks with restricted routes. In R. Din-
gledine, editor, Privacy Enhancing Technologies (PET 2003).
Springer-Verlag LNCS 2760, 2003.
[14] G. Danezis. Statistical disclosure attacks. In Security and
Privacy in the Age of Uncertainty (SEC2003), pages 421–426,
Athens, May 2003. IFIP TC11, Kluwer.
[15] G. Danezis, R. Dingledine, and N. Mathewson. Mixminion:
Design of a type III anonymous remailer protocol. In 2003
IEEE Symposium on Security and Privacy, pages 2–15. IEEE
CS, May 2003.
[16] D. Dean and A. Stubbleﬁeld. Using Client Puzzles to Protect
TLS. In Proceedings of the 10th USENIX Security Symposium.
USENIX, Aug. 2001.
[17] T. Dierks and C. Allen. The TLS Protocol Version 1.0.
IETF RFC 2246, January 1999.
[18] R. Dingledine, M. J. Freedman, D. Hopwood, and D. Molnar.
A Reputation System to Increase MIX-net Reliability. In I. S.
Moskowitz, editor, Information Hiding (IH 2001), pages 126–
141. Springer-Verlag, LNCS 2137, 2001.
[19] R. Dingledine, M. J. Freedman, and D. Molnar. The free
haven project: Distributed anonymous storage service. In
H. Federrath, editor, Designing Privacy Enhancing Technolo-
gies: Workshop on Design Issue in Anonymity and Unobserv-
ability. Springer-Verlag, LNCS 2009, July 2000.
[20] R. Dingledine and N. Mathewson. Tor protocol speciﬁcations.
<http://freehaven.net/tor/tor-spec.txt>.
[21] R. Dingledine and P. Syverson. Reliable MIX Cascade Net-
works through Reputation. In M. Blaze, editor, Financial
Cryptography. Springer-Verlag, LNCS 2357, 2002.
[22] J. Douceur. The Sybil Attack. In Proceedings of the 1st Inter-
national Peer To Peer Systems Workshop (IPTPS), Mar. 2002.
[23] H. Federrath, A. Jerichow, and A. Pﬁtzmann. MIXes in mo-
bile communication systems: Location management with pri-
vacy. In R. Anderson, editor, Information Hiding, First Inter-
national Workshop, pages 121–135. Springer-Verlag, LNCS
1174, May 1996.
[24] M. J. Freedman and R. Morris. Tarzan: A peer-to-peer
anonymizing network layer. In 9th ACM Conference on Com-
puter and Communications Security (CCS 2002), Washington,
DC, November 2002.
[25] S. Goel, M. Robson, M. Polte, and E. G. Sirer. Herbivore: A
scalable and efﬁcient protocol for anonymous communication.
Technical Report TR2003-1890, Cornell University Comput-
ing and Information Science, February 2003.
[26] I. Goldberg. A Pseudonymous Communications Infrastructure
for the Internet. PhD thesis, UC Berkeley, Dec 2000.
[27] D. M. Goldschlag, M. G. Reed, and P. F. Syverson. Hiding
routing information. In R. Anderson, editor, Information Hid-
ing, First International Workshop, pages 137–150. Springer-
Verlag, LNCS 1174, May 1996.
[28] C. G
¨
ulc
¨
u and G. Tsudik. Mixing E-mail with Babel. In Net-
work and Distributed Security Symposium (NDSS 96), pages
2–16. IEEE, February 1996.
[29] A. Hintz. Fingerprinting websites using trafﬁc analysis. In
R. Dingledine and P. Syverson, editors, Privacy Enhancing
Technologies (PET 2002), pages 171–178. Springer-Verlag,
LNCS 2482, 2002.
[30] A. Jerichow, J. M
¨
uller, A. Pﬁtzmann, B. Pﬁtzmann, and
M. Waidner. Real-time mixes: A bandwidth-efﬁcient
anonymity protocol. IEEE Journal on Selected Areas in Com-
munications, 16(4):495–509, May 1998.
[31] D. Kesdogan, D. Agrawal, and S. Penz. Limits of anonymity
in open environments. In F. Petitcolas, editor, Information
Hiding Workshop (IH 2002). Springer-Verlag, LNCS 2578,
October 2002.
[32] D. Koblas and M. R. Koblas. SOCKS. In UNIX Security III
Symposium (1992 USENIX Security Symposium), pages 77–
83. USENIX, 1992.
[33] B. N. Levine, M. K. Reiter, C. Wang, and M. Wright. Timing
analysis in low-latency mix-based systems. In A. Juels, ed-
itor, Financial Cryptography. Springer-Verlag, LNCS (forth-
coming), 2004.
[34] B. N. Levine and C. Shields. Hordes: A multicast-based proto-
col for anonymity. Journal of Computer Security, 10(3):213–
240, 2002.
[35] C. Meadows. The NRL protocol analyzer: An overview. Jour-
nal of Logic Programming, 26(2):113–131, 1996.
[36] U. M
¨
oller, L. Cottrell, P. Palfrader, and L. Sassaman. Mix-
master Protocol Version 2. Draft, July 2003. <http:
//www.abditum.com/mixmaster-spec.txt>.
[37] V. S. Pai, L. Wang, K. Park, R. Pang, and L. Peterson. The
Dark Side of the Web: An Open Proxy’s View.
<http://codeen.cs.princeton.edu/>.
[38] A. Pﬁtzmann, B. Pﬁtzmann, and M. Waidner. ISDN-mixes:
Untraceable communication with very small bandwidth over-
head. In GI/ITG Conference on Communication in Distributed
Systems, pages 451–463, February 1991.
[39] Privoxy. <http://www.privoxy.org/>.
[40] M. G. Reed, P. F. Syverson, and D. M. Goldschlag. Proto-
cols using anonymous connections: Mobile applications. In
B. Christianson, B. Crispo, M. Lomas, and M. Roe, editors,
Security Protocols: 5th International Workshop, pages 13–23.
Springer-Verlag, LNCS 1361, April 1997.
[41] M. G. Reed, P. F. Syverson, and D. M. Goldschlag. Anony-
mous connections and onion routing. IEEE Journal on Se-
lected Areas in Communications, 16(4):482–494, May 1998.
[42] M. K. Reiter and A. D. Rubin. Crowds: Anonymity for web
transactions. ACM TISSEC, 1(1):66–92, June 1998.
[43] M. Rennhard and B. Plattner. Practical anonymity for the
masses with morphmix. In A. Juels, editor, Financial Cryp-
tography. Springer-Verlag, LNCS (forthcoming), 2004.
[44] M. Rennhard, S. Rafaeli, L. Mathy, B. Plattner, and D. Hutchi-
son. Analysis of an Anonymity Network for Web Browsing.
In IEEE 7th Intl. Workshop on Enterprise Security (WET ICE
2002), Pittsburgh, USA, June 2002.
[45] A. Serjantov and P. Sewell. Passive attack analysis for
connection-based anonymity systems. In Computer Security
ESORICS 2003. Springer-Verlag, LNCS 2808, October 2003.
[46] R. Sherwood, B. Bhattacharjee, and A. Srinivasan. p
5
: A pro-
tocol for scalable anonymous communication. In IEEE Sym-
posium on Security and Privacy, pages 58–70. IEEE CS, 2002.
[47] A. Shubina and S. Smith. Using caching for browsing
anonymity. ACM SIGEcom Exchanges, 4(2), Sept 2003.
[48] P. Syverson, M. Reed, and D. Goldschlag. Onion Routing
access conﬁgurations. In DARPA Information Survivability
Conference and Exposition (DISCEX 2000), volume 1, pages
34–40. IEEE CS Press, 2000.
[49] P. Syverson, G. Tsudik, M. Reed, and C. Landwehr. Towards
an Analysis of Onion Routing Security. In H. Federrath, ed-
itor, Designing Privacy Enhancing Technologies: Workshop
on Design Issue in Anonymity and Unobservability, pages 96–
114. Springer-Verlag, LNCS 2009, July 2000.
[50] A. Tannenbaum. Computer networks, 1996.
[51] The AN.ON Project. German police proceeds against
anonymity service. Press release, September 2003.
<http://www.datenschutzzentrum.de/
material/themen/presse/anon-bka_e.htm>.
[52] M. Waldman and D. Mazi

eres. Tangler: A censorship-
resistant publishing system based on document entangle-
ments. In 8
th
ACM Conference on Computer and Commu-
nications Security (CCS-8), pages 86–135. ACM Press, 2001.
[53] M. Waldman, A. Rubin, and L. Cranor. Publius: A robust,
tamper-evident, censorship-resistant and source-anonymous
web publishing system. In Proc. 9th USENIX Security Sym-
posium, pages 59–72, August 2000.
[54] M. Wright, M. Adler, B. N. Levine, and C. Shields. Defending
anonymous communication against passive logging attacks. In
IEEE Symposium on Security and Privacy, pages 28–41. IEEE
CS, May 2003.
This used to be a fairly common mistake for Tor users early on. This would happen when people would not properly configure their applications when using Tor. With the advent of things like the Tor browser bundle which comes properly configured and ready to use out of the box, these issues have become less and less common. Tor is not a completely decentralized peer-to-peer system like Bitcoin for instance. Tor still requires a set of directory servers that manage and keep the state of the network. *Perfect forward secrecy* essentially means that future compromises of secret keys do not compromise past sessions of encrypted communication. This can be of utmost importance in Tor. Imagine some entity legally enforces the people running a set of Tor nodes to reveal all the secret keys being used by those nodes at the present time. In that scenario, even tho the anonymity of communications going through those nodes at that instance might affected, past sessions are not compromised. End-to-end traffic timing attacks typically involve an entity with control of both the first and last node of a path. With access to the traffic coming in and out of the path that entity will attempt to map requests coming into the first node to requests coming out of the last node of a circuit by looking at traffic patterns. The less traffic is going thru that path the easier it becomes to successfully perform traffic timing attacks. Researchers always guessed that government intelligence agencies like the NSA and its British equivalent GCHQ kept a close eye on anonymity systems like Tor. Edward Snowden leaked a number of documents that confirmed exactly that. The guardian published the [Tor Stinks Presentation](http://www.theguardian.com/world/interactive/2013/oct/04/tor-stinks-nsa-presentation-document) and more documents can be found at [https://edwardsnowden.com/](https://edwardsnowden.com/) This is a very important point to take into consideration when thinking about anonymity. [In December of 2013 during the Final Exams period at Harvard a student decided to try to delay a Final Exam by sending in a fake bomb threat](http://www.thecrimson.com/article/2013/12/17/student-charged-bomb-threat/). The student, Eldo Kim, took the steps to disguise his identity. He used [Guerrilla Mail](https://www.guerrillamail.com/) (a service which allows users to create free, disposable email addresses) to send the bomb threat to school officials and he accessed that service via Tor. Even though he took such precautions he still got caught. Here’s why: Turns out Guerrilla Mail sends a X-Originating-IP header along with the email message that specifies the IP of the computer that accessed Guerrilla Mail. Eldo used Tor to access Guerrilla Mail, which meant that the X-Originating-IP` was actually the IP of a Tor exit node (how do you know this by looking at the IP? remember that the IPs of most Tor nodes are readily available in the directory servers). Once the authorities knew that the request came from Tor they went and checked the IPs within the Harvard network that had accessed Tor at the time the email was sent. Because Harvard’s network uses a MAC address-based authorization system (you have to register your computer and link it with your Harvard ID in order to access the network) they were able to identify the few people accessing Tor via the Harvard network at the time the bomb threat came in. After that, the authorities went and interviewed those people, one of which was Aldo, who ended up admitting to sending the bomb threat. Aldo’s mistake was to overlook the fact that Tor does not try to conceal who is connected to the network. In the end, he could have denied the whole thing when the police came to talk to him and just claimed to have connected to Tor for some other reason. Nevertheless, there are a few other things he could have done: - Connected to Tor via a bridge node - Used a proxy to connect to Tor - Connected to the internet via the wi-fi at a cafe or some other public space Tor’s anonymity goals are two-folded: - Tor wants to provide anonymity for clients who which to connect to servers on the internet. This is called initiator anonymity. *e.g. Alice wants to visit a website that sells flowers without other people knowing* - Tor wants to provide anonymity for servers who want to make a service available on the internet. This is called responder anonymity and it is usually harder to ensure than initiator anonymity. *e.g. Alice wants to anonymously set up a service that sells flowers online*. Responder anonymity is what hidden services are used for. A well known real world example was [Silk Road](https://en.wikipedia.org/wiki/Silk_Road_(marketplace)). When you run a Tor exit node, that computer will be executing requests on behalf of anonymous users. Therefore, there is a chance your computer might, for instance, make HTTP requests to fetch or post illegal content e.g. child pornography (which means you could be confused with the person actually accessing the illegal content). Instances of people getting into legal troubles due to this type of reason are fairly rare. That said, you might want to be extra careful depending on where you live and who is your ISP. Exit policies, which essentially define a set of addresses and ports you are willing to connect to, were put in place to mitigate this type of issues. It is hard to know exactly what type of tactics powerful attackers (such as government agencies) could be employing to try to deanonymize Tor traffic. There is a chance that they have figured out some way of cracking Tor’s anonymity. Nevertheless, if we take a look at a number of the high profile cases in which the authorities were able to figure out the identity of a Tor user, we can usually find some compelling evidence of human error that could have indeed provided the authorities with enough clues as to be able to deduce the identity of the Tor user. A couple of examples: - *[The Harvard bomb threat](http://www.thecrimson.com/article/2013/12/17/student-charged-bomb-threat/)* - user did not use a proxy to connect to Tor (I go into detail about this case in a note on page 4 of the paper) - *[Silk Road](http://www.wired.com/2015/05/silk-road-creator-ross-ulbricht-sentenced-life-prison/)* - There were a string of mistakes Ross Ulbricht (the founder of Silk Road) made, such as linking his real identity with his online nickname and then writing posts related to Silk Road in forums using his real identity. Obviously, if government agencies did have a way to deanonymize Tor traffic they would likely still try to convince the general public they don’t. To accomplish this they would have to provide logical explanations as to how they would be able to crack cases without the secret key to Tor. Tor needs to have a substantial set of users in order for it to have strong anonymity guarantees and to be useful for real world use. Therefore, when facing design decisions you have to be very mindful of choices that might affect the adoption and popularity of Tor. If in order to run Tor you need to modify the Kernel of your Operating System, chances are a lot less people would be able to run Tor nodes which would in turn make the system more fragile. The notion of **symmetric-key** comes from **Symmetric-key encryption algorithms** (in contrast to Public Key Cryptography, look below). In Symmetric-key algorithms, the same key is used for encryption and decryption. This means that in order to exchange messages encrypted using Symmetric-key algorithms, the two parties have to first establish a secret key they both know about. If a third party somehow learns about this secret key, he will be able to decrypt messages. It is useful to note that Symmetric Key encryption is usually much less computationally expensive than Public Key encryption. I will also take the opportunity to quickly talk about *Public Key Cryptography* and *Digital signatures* as these topics are important to understand this paper. **Public Key Cryptography** Public-key cryptography is a set of cryptographic protocols based on algorithms that require two separate keys: - Private-key - which as the name indicates is meant to be secret - Public-key - which is public / visible to others These two keys are mathematically linked. In public-key cryptography the public key is used to encrypt plaintext, where the private key is used to decrypt cipher text. This means that any person can encrypt a message for a specific user, e.g. Alice, using Alice’s public key, which is known to everyone. However, the message can only be decrypted using Alice’s private key, which only Alice has access to. **Digital Signatures** Digital signatures make heavy use of public-key cryptography. You can think of a digital signature as somewhat similar to a physical signature. A digital signature is also used to prove the authenticity of a document/digital message. A digital signature binds an identity to a message. Only the person with the private key can produce valid signatures. Anybody with access to the public key can test the validity of the signatures. Say alice wants to digitally sign a message *m*. In order to do that Alice must have: - Private-key (signing key) - $KEY_{private}$ - Public-key (verification key) - $KEY_{public}$ Alice then uses the *signing* function to produce a valid signature: $$signing(message, KEY_{private}) \rightarrow signature$$ Don’t worry about the internals of the *signing* function. What you need to know is that it takes a *message* and the Private-key and it will produce a signature (a short string). Again, only a person who possesses a private key can produce a valid signature. Anyone can use the public key to verify the signature: $$verify(m,signature,KEY_{public}) \rightarrow true\ or\ false$$ The pre-alpha version of Tor was launched in September of 2002 ([here is the original email to the Freehaven mailing list](http://archives.seul.org/or/dev/Sep-2002/msg00019.html)). The paper you are reading was presented at the USENIX Security Symposium in 2004, and in 2006 Roger Dingledine, Nick Mathewson and five others founded [The Tor Project](https://www.torproject.org/), a Massachusetts-based nonprofit organization responsible for maintaining Tor. Tor is a computer system that enables anonymous communication. The core principle of Tor “onion routing” is a technique for anonymous communication over a public network. Onion Routing was developed in the mid-1990s at the U.S. Naval Research Laboratory ([here is one of the first papers about it](http://www.onion-router.net/Publications/JSAC-1998.pdf)). With onion routing messages are encapsulated in several layers of encryption, which are analogous to layers of an onion. The resulting “onion” is then transmitted through a series of nodes in a network (onion routers) with each node peeling a layer of encryption and therefore uncovering the data’s next destination. When the final layer is decrypted you get the original message (the plaintext). This way, the original author of the message remains anonymous because each intermediary in the network is only aware of the immediately preceding and following nodes in the path (except the first node that does know who the sender was, but does not know the final destination). Here is an illustration of Onion Routing: ![illustration](https://www.dropbox.com/s/6c7d3cl2tc48cmr/Screenshot%202016-02-01%2023.10.20.png?dl=1) It is important to note that the actual size of the onion does not increase as you add more layers of encryption (this is due to the nature of the encryption algorithms used in Onion Routing) Diffie–Hellman handshake (D–H) is a method of securely exchanging cryptographic keys over a public channel. The Diffie–Hellman handshake allows two parties that have not previously shared some trusted, secure channel of communication (e.g. being physically together or communicating via a trusted courier) to jointly establish a shared secret key over an insecure channel. Here is how it basically works: 1. We establish two prime numbers $g$ and $p$. 2. You pick a secret number $x$ which you keep to yourself. Then you compute $g^x \bmod p$ and send that result to me. 3. I pick a secret number $y$ and then send you $g^y \bmod p$ 4. You take $g^y \bmod p$ and compute $(g^y \bmod p)^x \bmod p$ 5. I do the equivalent and compute $(g^x \bmod p)^y \bmod p$ 6. And that’s it, now we share a secret key that only we know. That key is $(g^x \bmod p)^y \bmod p$. The trick is to realize that in the end we compute the same number. I compute: $$(g^x \bmod p)^y \bmod p = g^{xy} \bmod p$$ You compute: $$(g^y \bmod p)^x \bmod p = g^{yx} \bmod p$$ Those two are the same. Mixing, padding and traffic shaping are techniques that, in this context, could be used in order to try to protect against end-to-end attacks such as traffic timing attacks. Traffic timing attacks typically involve an entity with control of both the first and last node of a circuit, looking at traffic patterns to map requests coming into the first node to requests coming out of the last node of a circuit. The drawback of using these techniques is that they severely affect the usability of the system (e.g. would make accessing a website via Tor orders of magnitude slower) and as a consequence the designers of Tor decided not to use them.