482 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 16, NO. 4, MAY 1998
Anonymous Connections and Onion Routing
Michael G. Reed, Member, IEEE, Paul F. Syverson, and David M. Goldschlag
Abstract—Onion routing is an infrastructure for private com-
munication over a public network. It provides anonymous con-
nections that are strongly resistant to both eavesdropping and
traffic analysis. Onion routing’s anonymous connections are bidi-
rectional, near real-time, and can be used anywhere a socket
connection can be used. Any identifying information must be
in the data stream carried over an anonymous connection. An
onion is a data structure that is treated as the destination address
by onion routers; thus, it is used to establish an anonymous
connection. Onions themselves appear different to each onion
router as well as to network observers. The same goes for
data carried over the connections they establish. Proxy-aware
applications, such as web browsers and e-mail clients, require
no modification to use onion routing, and do so through a
series of proxies. A prototype onion routing network is running
between our lab and other sites. This paper describes anonymous
connections and their implementation using onion routing. This
paper also describes several application proxies for onion routing,
as well as configurations of onion routing networks.
Index Terms—Anonymity, communications, Internet, privacy,
security, traffic analysis.
I. INTRODUCTION
I
S INTERNET communication private? Most security con-
cerns focus on preventing eavesdropping,
1
i.e., outsiders
listening in on electronic conversations. But encrypted mes-
sages can still be tracked, revealing who is talking to whom.
This tracking is called traffic analysis and may reveal sensitive
information. For example, the existence of inter-company
collaboration may be confidential. Similarly, e-mail users
may not wish to reveal who they are communicating with
to the rest of the world. In certain cases anonymity may
be desirable, for example, anonymous e-cash is not very
anonymous if delivered with a return address. Web-based
shopping or browsing of public databases should not require
revealing one’s identity.
This paper describes how a freely available system,
onion routing, can be used to protect a variety of Internet
services against both eavesdropping and traffic analysis
attacks from both the network and outside observers. This
paper includes a specification sufficient to guide both re-
implementations and new applications of onion routing.
Manuscript received March, 1997; revised September, 1997. This work
was supported by the Office of Naval Research and by the Defense Advanced
Research Projects Agency.
M. G. Reed and P. F. Syverson are with the Naval Research Laboratory,
Center For High Assurance Computer Systems, Washington, D.C. 20375-
5337 USA (e-mail: reed@itd.nrl.navy.mil; syverson@itd.nrl.navy.mil).
D. M. Goldschlag is with Divx, Herndon, VA 20170 USA (e-mail:
david.goldschlag@divx.com).
Publisher Item Identifier S 0733-8716(98)01110-X.
1
Internet Engineering Task Force. HTTP://www. ietf. org.
We also discuss configurations of onion routing networks
and applications of onion routing, including virtual private
networks (VPN’s), Web browsing, e-mail, remote login, and
electronic cash.
2
One purpose of traffic analysis is to reveal who is talking
to whom. The anonymous connections described here are
designed to be resistant to traffic analysis, i.e., to make it
difficult for observers to learn identifying information from the
connection (e.g., by reading packet headers, tracking encrypted
payloads, etc.). Any identifying information must be passed as
data through the anonymous connections. Our implementation
of anonymous connections, onion routing, provides protection
against eavesdropping as a side effect. Onion routing provides
bidirectional and near real-time communication similar to
TCP/IP socket connections or ATM AAL5 [6]. The anonymous
connections can substitute for sockets in a wide variety of
unmodified Internet applications by means of proxies. Data
may also be passed through a privacy filter before being
sent over an anonymous connection. This removes identifying
information from the data stream, to make communication
anonymous too.
Although onion routing may be used for anonymous com-
munication, it differs from anonymous remailers [7], [15]
in two ways: communication is real-time and bidirectional,
and the anonymous connections are application independent.
Onion routing’s anonymous connections can support anony-
mous mail as well as other applications. For example, onion
routing may be used for anonymous Web browsing. A user
may wish to browse public Web sites without revealing his
identity to those Web sites. That requires removing infor-
mation that identifies him from his requests to Web servers
and removing information from the connection itself that
may identify him. Hence, anonymous Web browsing uses
anonymized communication over anonymous connections. The
Anonymizer [1] only anonymizes the data stream, not the
connection itself. So it does not prevent traffic analysis attacks
like tracking data as it moves through the network.
This paper is organized in the following way: Section II
presents an overview of onion routing. Section III presents
empirical data about our prototype. Section IV defines our
threat model. Section V describes onion routing and the appli-
cation specific proxies in more detail. Section VI describes the
implementation choices that were made for security reasons.
Section VII describes how onion routing may be used in a wide
variety of Internet applications. Section VIII contrasts onion
routing with related work, and Section IX presents concluding
remarks.
2
Preliminary versions of portions of this paper have appeared in [23], [13],
[19], and [14].
U.S. Government work not protected by U.S. copyright.
REED et al.: ANONYMOUS CONNECTIONS AND ONION ROUTING 483
II. ONION ROUTING OVERVIEW
In onion routing, instead of making socket connections
directly to a responding machine, initiating applications make
connections through a sequence of machines called onion
routers. The onion routing network allows the connection
between the initiator and responder to remain anonymous.
Anonymous connections hide who is connected to whom, and
for what purpose, from both outside eavesdroppers and com-
promised onion routers. If the initiator also wants to remain
anonymous to the responder, then all identifying information
must be removed from the data stream before being sent over
the anonymous connection.
Onion routers in the network are connected by longstand-
ing (permanent) socket connections. Anonymous connections
through the network are multiplexed over the longstanding
connections. For any anonymous connection, the sequence of
onion routers in a route is strictly defined at connection setup.
However, each onion router can only identify the previous
and next hop along a route. Data passed along the anonymous
connection appear different at each onion router, so data cannot
be tracked en route, and compromised onion routers cannot
cooperate by correlating the data stream each sees. We will
also see that they cannot make use of replayed onions or
replayed data.
A. Operational Overview
The onion routing network is accessed via a series of
proxies. An initiating application makes a socket connection to
an application proxy. This proxy massages connection message
format (and later data) to a generic form that can be passed
through the onion routing network. It then connects to an
onion proxy, which defines a route through the onion routing
network by constructing a layered data structure called an
onion. The onion is passed to the entry funnel, that occupies
one of the longstanding connections to an onion router and
multiplexes connections to the onion routing network at that
onion router. That onion router will be the one for whom the
outermost layer of the onion is intended. Each layer of the
onion defines the next hop in a route. An onion router that
receives an onion peels off its layer, identifies the next hop,
and sends the embedded onion to that onion router. The last
onion router forward data to an exit funnel, whose job is to pass
data between the onion routing network and the responder.
In addition to carrying next-hop information, each onion
layer contains key seed material from which keys are gen-
erated for crypting
3
data sent forward or backward along the
anonymous connection. (We define forward to be the direction
in which the onion travels and backward as the opposite
direction.)
Once the anonymous connection is established, it can carry
data. Before sending data over an anonymous connection,
the onion proxy adds a layer of encryption for each onion
router in the route. As data move through the anonymous
connection, each onion router removes one layer of encryption,
so it arrives at the responder as plaintext. This layering
3
We define the verb crypt to mean the application of a cryptographic
operation, be it encryption or decryption.
occurs in the reverse order for data moving back to the
initiator. Therefore data that have passed backward through
the anonymous connection must be repeatedly post-crypted to
obtain the plaintext.
By layering cryptographic operations in this way, we gain
an advantage over link encryption. As data move through the
network it appears different to each onion router. Therefore,
an anonymous connection is as strong as its strongest link,
and even one honest node is enough to maintain the privacy
of the route. In link encrypted systems, compromised nodes
can trivially cooperate to uncover route information.
Onion routers keep track of received onions until they
expire. Replayed or expired onions are not forwarded, so
they cannot be used to uncover route information, either by
outsiders or compromised onion routers. Note that clock skew
between onion routers can only cause an onion router to reject
a fresh onion or to keep track of processed onions longer than
necessary. Also, since data are encrypted using stream ciphers,
replayed data will look different each time it passes through
a properly operating onion router.
Although we call this system onion routing, the routing that
occurs here does so at the application layer of the protocol
stack and not at the IP layer. More specifically, we rely
upon IP routing to route data passed through the longstanding
socket connections. An anonymous connection is comprised
of portions of several linked longstanding multiplexed socket
connections. Therefore, although the series of onion routers
in an anonymous connection is fixed for the lifetime of that
anonymous connection, the route that data actually travels be-
tween individual onion routers is determined by the underlying
IP network. Thus, onion routing may be compared to loose
source routing.
Onion routing depends upon connection-based services that
deliver data uncorrupted and in order. This simplifies the
specification of the system. TCP socket connections, which are
layered on top of a connectionless service like IP, provide these
guarantees. Similarly, onion routing could easily be layered on
top of other connection based services, like ATM AAL5.
Our current prototype of onion routing considers the net-
work topology to be static and does not have mechanisms
to automatically distribute or update public keys or network
topology. These issues, though important, are not the key parts
of onion routing and will be addressed in a later prototype.
B. Configurations
As mentioned above, neighboring onion routers are neigh-
bors in virtue of having longstanding socket connections
between them, and the network as a whole is accessed from
the outside through a series of proxies. By adjusting where
those proxies reside it is possible to vary which elements
of the system are trusted by users and in what way. (For
some configurations it may be efficient to combine proxies that
reside in the same place, thus they may be only conceptually
distinct.)
1) Firewall Configuration: In the firewall configuration,an
onion router sits on the firewall of a sensitive site. This onion
router serves as an interface between machines behind the
484 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 16, NO. 4, MAY 1998
firewall and the external network. Connections from machines
behind the firewall to the onion router are protected by other
means (e.g., physical security). To complicate tracking of
traffic originating or terminating within the sensitive site,
this onion router should also route data between other onion
routers. This configuration might represent the system interface
from a typical corporate or government site. Here the applica-
tion proxies (together with any privacy filters) and the onion
proxies would typically live at the firewall as well. (Typically,
there might only be one onion proxy.)
There are three important features of this basic configura-
tion.
• Connections between machines behind onion routers are
protected against both eavesdropping and traffic analysis.
Since the data stream never appears in the clear on the
public network, this data may carry identifying informa-
tion but communication is still private. (This feature is
used in Section VII-A.)
• The onion router at the originally protected site knows
both the source and destination of a connection. This
protects the anonymity of connections from observers
outside the firewall but also simplifies enforcement of
and monitoring for compliance with corporate or govern-
mental usage policy.
• The use of anonymous connections between two sensitive
sites that both control onion routers effectively hides their
communication from outsiders. However, if the responder
is not in a sensitive site (e.g., the responder is some
arbitrary Web server), the data stream from the sensitive
initiator must also be anonymized. If the connection
between the exit funnel and the responding server is
unencrypted, the data stream might otherwise identify the
initiator. For example, an attacker could simply listen in
on the connections to a Web server and identify initiators
of any connection to it.
2) Remote Proxy Configuration: What happens if an ini-
tiator does not control an onion router? If the initiator can
make encrypted connections to some remote onion router, then
he can function as if he is in the firewall configuration just
described, except that both observers and the network can tell
when he makes connections to the onion router. However,
if the initiator trusts the onion router to build onions, his
association with the anonymous connection from that onion
router to the responder is hidden from observers and the
network. In a similar way, an encrypted connection from
an exit funnel to a responder hides the association of the
responder with the anonymous connection.
Therefore, if an initiator makes an anonymous connection
to some responder, and layers end-to-end encryption over that
anonymous connection, the initiator and responder can identify
themselves to one another, yet hide their communication from
the rest of the world.
Notice, however, that the initiator trusts the remote onion
router to conceal that the initiator wants to communicate with
the responder and to build an anonymous connection through
other onion routers. The next section describes how to shift
some of this trust from the first onion router to the initiator.
3) The Customer-ISP Configuration: Suppose, for exam-
ple, an Internet Services Provider (ISP) runs a funnel
that accepts connections from onion proxies running on
subscribers’ machines. In this configuration, users generate
onions specifying a path through the ISP to the destination.
Although the ISP would know who initiates the connection,
the ISP would not know with whom the customer is
communicating, nor would it be able to see data content.
So the customer need not trust the ISP to maintain her
privacy. Furthermore, the ISP becomes a common carrier
that carries data for its customers. This may relieve the ISP of
responsibility for both whom users are communicating with
and the content of those conversations. The ISP may or may
not be running an onion router as well. If he is running an
onion router, then it is more difficult to identify connections
that terminate with his customers; however, he is serving as a
routing point for other traffic. On the other hand, if he simply
runs a funnel to an onion router elsewhere, it will be possible
to identify connections terminating with him, but his overall
traffic load will be less. Which of these would be the case for a
given ISP would probably depend on a variety of service, cost,
and pricing considerations. Note that in this configuration, the
entry funnel must have an established longstanding connection
to an onion router just like any neighboring onion router (cf.
Section V-F for a description of how these are established).
But, in most other cases, where the funnel resides on the
same machine as the onion router, establishing an encrypted
longstanding connection should not be necessary because the
funnel can be directly incorporated into the onion router.
III. E
MPIRICAL DATA
We invite readers to experiment with our prototype of
onion routing by using it to anonymously surf the Web, send
anonymous e-mail, and do remote logins. For instructions,
please see http://www.onion.router.net/ .
One should be aware that accessing a remote onion router
does not completely preserve anonymity because the connec-
tion between a remote machine and the first onion router is
not protected. If that connection were protected, one would be
in the remote proxy configuration, but there would still be no
reason to trust the remote onion router. If one had a secured
connection to an onion router one trusted, our onion router
could be used as one of several intermediate routers to further
complicate traffic analysis.
We have recently set up a 13-node distributed network of
government, academic, and private sites. However, at press
time we have not yet gathered performance data for this
network. The data we present are for a network running on
a single machine. In our experimental onion routing network,
five onion routers run on a single Sun Ultra 2 2170. This
machine has two 167-MHz processors and 256 MB of mem-
ory. Anonymous connections are routed through a random
sequence of five onion routers. Connection setup time should
be comparable to a more distributed topology. Data latency,
however, is more difficult to judge. Clearly, data will travel
faster over socket connections between onion routers on the
same machine than over socket connections between different
REED et al.: ANONYMOUS CONNECTIONS AND ONION ROUTING 485
machines. However, on a single machine the removal or
addition of layers of encryption is not pipelined, so data latency
may be worse.
Onion routing’s overhead is mainly due to public key
cryptography and is incurred while setting up an anonymous
connection. On our Ultra 2, running a fast implementation of
RSA [2], a single public key decryption of a 1024-b plaintext
block using a 1024-b private key and a 1024-b modulus, takes
90 ms. Encryption is much faster, because the public keys
are only 16 b long. (This is why RSA signature verification
is cheaper than signing.) So, the public key cryptographic
overhead for routes spanning five onion routers is just under
0.5 s. This overhead can be further reduced, either with
specialized hardware, or even simply on different hardware
(a 200-MHz Pentium would be almost twice as fast).
In practice, our connection setup overhead does not appear
to add intolerably to the overhead of typical socket connec-
tions. Still, it can be further reduced. There is no reason that
the same anonymous connection could not be used to carry the
traffic for several “real” socket connections, either sequentially
or multiplexed. In fact, the specification for HTTP 1.1 defines
pipelined connections to amortize the cost of socket setup,
and pipelined connections would also transparently amortize
the increased cost of anonymous connection setup. We are
currently updating our Web proxy to be HTTP 1.1 compliant.
IV. T
HREAT MODEL
This section outlines our threat model. It does not intend
to quantify the cost of attacks, but to define possible attacks.
Future work will quantify the threat. First, some vocabulary. A
session is the data carried over a single anonymous connection.
Data are carried in fixed length cells. Because these cells
are multiply encrypted and change as they move through an
anonymous connection, tracking cells is equivalent to tracking
markers that indicate when cells begin. In a marker attack,
the attacker identifies the set of outbound connections that
some distinguished marker may have been forwarded upon.
By intersecting these sets for a series of distinguished markers
belonging to the same session, an attacker may determine, or
at least narrow, the set of possible next hops. In a timing
attack, the attacker records a timing signature for a session
that correlates data rate over time. A session may have a
very similar timing signature wherever it is measured over
a route, so cooperating attackers may determine if they carry
a particular session.
We assume that the network is subject to both passive and
active attacks. Traffic may be monitored and modified by both
external observers and internal network elements, including
compromised onion routers. Attackers may cooperate and
share information and inferences. We assume roving attackers
that can monitor part, but not all, of the network at a time.
Our goal is to prevent traffic analysis, not traffic confirma-
tion. If an attacker wants to confirm that two endpoints often
communicate, and he observes that they each connect to an
anonymous connection at roughly the same time, more often
than is statistically expected, it is reasonable to infer that the
endpoints are indeed communicating. Notice that this attack
is infeasible if endpoints live in protected networks behind
trusted onion routers on firewalls.
If the onion routing infrastructure is uniformly busy, then
passive external attacks are ineffective. Specifically, neither the
marker nor timing attacks are feasible, since external observers
cannot assign markers to sessions. Active attacks are possible,
because reducing the load on the system makes the network
easier to analyze (and makes the system not uniformly busy).
Passive internal attacks require at least two compromised
onion routers. Since onion routers can assign markers to a
session, both the marker and timing attacks are possible.
Specifically, timing signatures can be broadcast, and other
compromised onion routers can attempt to find connections
with matching timing signatures.
Another attack that is only feasible as an internal attack
is the volume attack. Compromised onion routers can keep
track of the number of cells that have passed over any given
anonymous connection. They can then simply broadcast totals
to other compromised onion routers. Cell totals that are close
to the same amount at the same time at different onion routers
are likely to belong to the same anonymous connection.
4
Active internal attacks amplify these risks, since individual
onion routers can selectively limit traffic on particular connec-
tions. An onion router, for example, could force a particular
timing signature on a connection and advertise that signature.
V. O
NION ROUTING SPECIFICS
A. Onion Routing Proxies
A proxy is a transparent service between two applications
that would usually make a direct socket connection to each
other but cannot. For example, a firewall might prevent direct
socket connections between internal and external machines. A
proxy running on the firewall may enable such connections.
Proxy-aware applications are becoming quite common.
Our goal has been to design an architecture for private
communication that would interface with unmodified appli-
cations, so we chose to use proxies as the interface between
applications and onion routing’s anonymous connections. For
applications that are designed to be proxy aware (e.g., WWW
browsers), we simply design appropriate interface proxies.
Surprisingly, for certain applications that are not proxy aware
(e.g., RLOGIN), we have also been able to design interface
proxies.
Because it is necessary to bridge between applications
and the onion routing network, proxies must understand
both application protocols and onion routing protocols.
Therefore, we modularize the design into components: the
application proxy, the onion proxy, and the entry funnel.
The application proxy bridges between a socket connection
from an application and a socket connection to the onion
proxy. It is the obligation of the application proxy to
massage the data stream so the onion proxy, the entry
funnel, and the exit funnel can be application independent.
Specifically, the application proxy must prepend to the
data stream a standard structure that identifies the ultimate
4
Thanks to Gene Tsudik for noting this attack and for helpful discussions.
486 IEEE JOURNAL ON SELECTED AREAS IN COMMUNICATIONS, VOL. 16, NO. 4, MAY 1998
destination by either hostname/port or IP address/port.
Additionally, it must process a 1-byte return code from
the exit funnel and either continue if no error is reported
or report the onion routing error code in some application
specific meaningful way. The application proxy may also
contain an optional privacy filter for sanitizing the data
stream.
Upon receiving a new request, the onion proxy builds an
onion defining the route of an anonymous connection. (It
may use the destination address in the prepended structure
to help define the route.) It then passes the onion to the
funnel, and repeatedly precrypts the standard structure. Finally,
it passes the precrypted standard structure through the anony-
mous connection to the exit funnel, thus specifying the ultimate
destination. From this point on, the onion proxy blindly relays
data back and forth between the application proxy and the
onion routing network (and thus the exit funnel at the other
end of the anonymous connection). Of course, it must apply the
appropriate keystreams to incoming and outgoing data when
blindly relaying data.
The entry funnel multiplexes connections from onion prox-
ies to the onion routing network. For the services we have
considered to date, a nearly generic exit funnel is adequate.
Its function is to demultiplex connections from the last onion
router to the outside. When it reads a data stream from the
terminating onion router, the first datum received will be the
standard structure specifying the ultimate destination. The exit
funnel makes a socket connection to that IP address/port,
reports a one-byte status message back to the onion routing
network (and thus back to the onion proxy which in turn
forward it back to the application proxy), and subsequently
moves data between the onion routing network and the new
socket. (For certain services, like RLOGIN, the exit funnel
also infers that the new socket must originate from a trusted
port.) Entry and exit funnels are not application specific but
must understand the onion routing protocol, that defines how
multiplexed connections are handled.
As an example, consider the application proxy for HTTP.
The user configures his browser to use the onion routing
proxy. His browser may send the proxy a request like
GET
http://www.onion-router.net/index.html
HTTP1.0 followed by optional fields.
The application proxy is listening for new requests. Once
it obtains the GET request, it creates the standard structure
and sends it (along a new socket connection) to the onion
proxy, to inform the onion proxy of the service and destination
of the anonymous connection. The application proxy then
modifies the GET request to GET/index.html HTTP/1.0
and sends it directly (through the anonymous connection) to
the HTTP server www.onion-router.net, followed by
the optional fields. Notice that the server name and http://
are eliminated from the GET request because the connection
is made directly to the HTTP server.
The application proxy essentially makes a connection to
www.onion-router.net, and issues a request as if it were
a client. Once this request is transmitted to the server, all
proxies blindly forward data in both directions between the
client and the server until the socket is broken by either side.
Fig. 1. The standard structure.
For the anonymizing onion routing HTTP proxy, the appli-
cation proxy proceeds as outlined above with one change: It
is now necessary to sanitize the optional fields that follow the
GET command because they may contain identity information.
Furthermore, the data stream during a connection must be
monitored to sanitize additional headers that might occur
during the connection. For our current anonymizing HTTP
proxy, operations that store cookies on the user’s browser (to
track a user, for example) are removed. This reduces function,
so applications that depend upon cookies (like online shopping
baskets) may not work properly.
B. Implementation
This section presents the interface specification between the
components in an onion routing system. To provide some
structure to this specification, we will discuss components in
the order that data would move from an initiating client to a
responding server.
There are four phases in an onion routing system: network
setup, that establishes the longstanding connections between
onion routers; connection setup, which establishes anonymous
connections through the onion router network; data move-
ment over an anonymous connection; and the destruction and
cleanup of anonymous connections. We will commingle the
discussion of these below.
C. Application Proxy
The interface between an application and the application
proxy is application specific. The interface between the appli-
cation proxy and the onion proxy is defined as follows: For
each new proxy request, the application proxy first determines
if it will handle or deny the request. If rejected, it reports an
application-specific error message and then closes the socket
and waits for the next request. If accepted, it creates a socket
to the onion proxy’s well-known port. The application proxy
then sends a standard structure to the onion proxy of the form
as shown in Fig. 1.
Version is currently defined to be 1. Protocol is either 1
for RLOGIN, 2 for HTTP, or 3 for SMTP. Retry Count
specifies how many times the exit funnel should attempt to
retry connecting to the ultimate destination. Finally, the Addr
Format field specifies the form of the ultimate destination
address: 1 for a NULL terminated ASCII string with the
hostname or IP address (in ASCII form) immediately followed
by another NULL terminated ASCII string with the destination
port number, and all others currently undefined. The ultimate
destination address is sent after this standard structure, and
the application proxy waits for a one byte error code before
sending data.