| blob | 5fac644e6225ba2c2201203144b157df5ad0b642 |
4 <head>
18 </head>
19 <body>
34 We present Tor, a circuit-based low-latency anonymous communication
35 service. This second-generation Onion Routing system addresses limitations
36 in the original design by adding perfect forward secrecy, congestion
37 control, directory servers, integrity checking, configurable exit policies,
38 and a practical design for location-hidden services via rendezvous
39 points. Tor works on the real-world
40 Internet, requires no special privileges or kernel modifications, requires
41 little synchronization or coordination between nodes, and provides a
42 reasonable tradeoff between anonymity, usability, and efficiency.
43 We briefly describe our experiences with an international network of
44 more than 30 nodes. We close with a list of open problems in anonymous communication.
54 </a>
57 Onion Routing is a distributed overlay network designed to anonymize
58 TCP-based applications like web browsing, secure shell,
59 and instant messaging. Clients choose a path through the network and
61 in the path knows its predecessor and successor, but no other nodes in
62 the circuit. Traffic flows down the circuit in fixed-size
64 (like the layers of an onion) and relayed downstream. The
65 Onion Routing project published several design and analysis
66 papers [<a href="#or-ih96" name="CITEor-ih96">27</a>,<a href="#or-jsac98" name="CITEor-jsac98">41</a>,<a href="#or-discex00" name="CITEor-discex00">48</a>,<a href="#or-pet00" name="CITEor-pet00">49</a>]. While a wide area Onion
67 Routing network was deployed briefly, the only long-running
68 public implementation was a fragile
69 proof-of-concept that ran on a single machine. Even this simple deployment
70 processed connections from over sixty thousand distinct IP addresses from
71 all over the world at a rate of about fifty thousand per day.
72 But many critical design and deployment issues were never
73 resolved, and the design has not been updated in years. Here
74 we describe Tor, a protocol for asynchronous, loosely federated onion
75 routers that provides the following improvements over the old Onion
76 Routing design:
80 a single hostile node could record traffic and
81 later compromise successive nodes in the circuit and force them
82 to decrypt it. Rather than using a single multiply encrypted data
85 where the initiator negotiates session keys with each successive hop in
86 the circuit. Once these keys are deleted, subsequently compromised nodes
87 cannot decrypt old traffic. As a side benefit, onion replay detection
88 is no longer necessary, and the process of building circuits is more
89 reliable, since the initiator knows when a hop fails and can then try
90 extending to a new node.
94 Onion Routing originally required a separate "application
96 never written, so many applications were never supported. Tor uses the
97 standard and near-ubiquitous SOCKS [<a href="#socks4" name="CITEsocks4">32</a>] proxy interface, allowing
98 us to support most TCP-based programs without modification. Tor now
99 relies on the filtering features of privacy-enhancing
100 application-level proxies such as Privoxy [<a href="#privoxy" name="CITEprivoxy">39</a>], without trying
101 to duplicate those features itself.
105 Routing originally called for batching and reordering cells as they arrived,
106 assumed padding between ORs, and in
107 later designs added padding between onion proxies (users) and
108 ORs [<a href="#or-ih96" name="CITEor-ih96">27</a>,<a href="#or-jsac98" name="CITEor-jsac98">41</a>]. Tradeoffs between padding protection
110 theorized [<a href="#or-pet00" name="CITEor-pet00">49</a>] to provide good security without expensive
111 padding, but no concrete padding scheme was suggested.
113 and deployment experience [<a href="#freedom21-security" name="CITEfreedom21-security">4</a>] suggest that this
114 level of resource use is not practical or economical; and even full
115 link padding is still vulnerable [<a href="#defensive-dropping" name="CITEdefensive-dropping">33</a>]. Thus,
116 until we have a proven and convenient design for traffic shaping or
117 low-latency mixing that improves anonymity against a realistic
118 adversary, we leave these strategies out.
122 built a separate circuit for each
123 application-level request, but this required
124 multiple public key operations for every request, and also presented
125 a threat to anonymity from building so many circuits; see
127 streams along each circuit to improve efficiency and anonymity.
131 within the circuit, Tor initiators can direct traffic to nodes partway
132 down the circuit. This novel approach
133 allows traffic to exit the circuit from the middle — possibly
134 frustrating traffic shape and volume attacks based on observing the end
135 of the circuit. (It also allows for long-range padding if
136 future research shows this to be worthwhile.)
140 address traffic bottlenecks. Unfortunately, typical approaches to
141 load balancing and flow control in overlay networks involve inter-node
142 control communication and global views of traffic. Tor's decentralized
143 congestion control uses end-to-end acks to maintain anonymity
144 while allowing nodes at the edges of the network to detect congestion
145 or flooding and send less data until the congestion subsides.
149 planned to flood state information through the network — an approach
150 that can be unreliable and complex. Tor takes a simplified view toward distributing this
151 information. Certain more trusted nodes act as <em>directory
152 servers</em>: they provide signed directories describing known
153 routers and their current state. Users periodically download them
154 via HTTP.
158 for each node to advertise a policy describing the hosts
159 and ports to which it will connect. These exit policies are critical
160 in a volunteer-based distributed infrastructure, because each operator
161 is comfortable with allowing different types of traffic to exit
162 from his node.
166 design did no integrity checking on data. Any node on the
167 circuit could change the contents of data cells as they passed by — for
168 example, to alter a connection request so it would connect
169 to a different webserver, or to `tag' encrypted traffic and look for
170 corresponding corrupted traffic at the network edges [<a href="#minion-design" name="CITEminion-design">15</a>].
171 Tor hampers these attacks by verifying data integrity before it leaves
172 the network.
178 Tor provides an integrated mechanism for responder anonymity via
179 location-protected servers. Previous Onion Routing designs included
180 long-lived "reply onions" that could be used to build circuits
181 to a hidden server, but these reply onions did not provide forward
182 security, and became useless if any node in the path went down
184 to connect with hidden servers; reply onions are no longer required.
187 Unlike Freedom [<a href="#freedom2-arch" name="CITEfreedom2-arch">8</a>], Tor does not require OS kernel
188 patches or network stack support. This prevents us from anonymizing
189 non-TCP protocols, but has greatly helped our portability and
190 deployability.
195 We have implemented all of the above features, including rendezvous
196 points. Our source code is
197 available under a free license, and Tor
198 is not covered by the patent that affected distribution and use of
199 earlier versions of Onion Routing.
200 We have deployed a wide-area alpha network
201 to test the design, to get more experience with usability
202 and users, and to provide a research platform for experimentation.
203 As of this writing, the network stands at 32 nodes spread over two continents.
208 and then address the above list of improvements in
209 Sections <a href="#sec:design">4</a>, <a href="#sec:rendezvous">5</a>, and <a href="#sec:other-design">6</a>.
210 We summarize
212 known attacks, and talk about our early deployment experiences in
223 </a>
228 proposed hiding the correspondence between sender and recipient by
229 wrapping messages in layers of public-key cryptography, and relaying them
230 through a path composed of "mixes." Each mix in turn
231 decrypts, delays, and re-orders messages before relaying them
232 onward.
235 Subsequent relay-based anonymity designs have diverged in two
239 to maximize anonymity at the cost of introducing comparatively large and
241 networks resist strong global adversaries,
242 but introduce too much lag for interactive tasks like web browsing,
243 Internet chat, or SSH connections.
247 try to anonymize interactive network traffic. These systems handle
248 a variety of bidirectional protocols. They also provide more convenient
249 mail delivery than the high-latency anonymous email
250 networks, because the remote mail server provides explicit and timely
251 delivery confirmation. But because these designs typically
252 involve many packets that must be delivered quickly, it is
253 difficult for them to prevent an attacker who can eavesdrop both ends of the
254 communication from correlating the timing and volume
255 of traffic entering the anonymity network with traffic leaving it [<a href="#SS03" name="CITESS03">45</a>].
256 These
257 protocols are similarly vulnerable to an active adversary who introduces
258 timing patterns into traffic entering the network and looks
259 for correlated patterns among exiting traffic.
260 Although some work has been done to frustrate these attacks, most designs
261 protect primarily against traffic analysis rather than traffic
265 The simplest low-latency designs are single-hop proxies such as the
266 <b>Anonymizer</b> [<a href="#anonymizer" name="CITEanonymizer">3</a>]: a single trusted server strips the
267 data's origin before relaying it. These designs are easy to
268 analyze, but users must trust the anonymizing proxy.
269 Concentrating the traffic to this single point increases the anonymity set
270 (the people a given user is hiding among), but it is vulnerable if the
271 adversary can observe all traffic entering and leaving the proxy.
274 More complex are distributed-trust, circuit-based anonymizing systems.
275 In these designs, a user establishes one or more medium-term bidirectional
276 end-to-end circuits, and tunnels data in fixed-size cells.
277 Establishing circuits is computationally expensive and typically
278 requires public-key
279 cryptography, whereas relaying cells is comparatively inexpensive and
280 typically requires only symmetric encryption.
281 Because a circuit crosses several servers, and each server only knows
282 the adjacent servers in the circuit, no single server can link a
283 user to her communication partners.
288 approach aggregates users into larger anonymity sets, but again an
289 attacker only needs to observe both ends of the cascade to bridge all
290 the system's traffic. The Java Anon Proxy's design
291 calls for padding between end users and the head of the
292 cascade [<a href="#web-mix" name="CITEweb-mix">7</a>]. However, it is not demonstrated whether the current
293 implementation's padding policy improves anonymity.
296 <b>PipeNet</b> [<a href="#back01" name="CITEback01">5</a>,<a href="#pipenet" name="CITEpipenet">12</a>], another low-latency design proposed
297 around the same time as Onion Routing, gave
298 stronger anonymity but allowed a single user to shut
299 down the network by not sending. Systems like <b>ISDN
300 mixes</b> [<a href="#isdn-mixes" name="CITEisdn-mixes">38</a>] were designed for other environments with
301 different assumptions.
304 In P2P designs like <b>Tarzan</b> [<a href="#tarzan:ccs02" name="CITEtarzan:ccs02">24</a>] and
305 <b>MorphMix</b> [<a href="#morphmix:fc04" name="CITEmorphmix:fc04">43</a>], all participants both generate
306 traffic and relay traffic for others. These systems aim to conceal
307 whether a given peer originated a request
308 or just relayed it from another peer. While Tarzan and MorphMix use
309 layered encryption as above, <b>Crowds</b> [<a href="#crowds-tissec" name="CITEcrowds-tissec">42</a>] simply assumes
310 an adversary who cannot observe the initiator: it uses no public-key
311 encryption, so any node on a circuit can read users' traffic.
314 <b>Hordes</b> [<a href="#hordes-jcs" name="CITEhordes-jcs">34</a>] is based on Crowds but also uses multicast
315 responses to hide the initiator. <b>Herbivore</b> [<a href="#herbivore" name="CITEherbivore">25</a>] and
316 <b>P</b><sup><b>5</b></sup> [<a href="#p5" name="CITEp5">46</a>] go even further, requiring broadcast.
317 These systems are designed primarily for communication among peers,
318 although Herbivore users can make external connections by
319 requesting a peer to serve as a proxy.
323 all at once, using a layered "onion" of public-key encrypted messages,
324 each layer of which provides session keys and the address of the
325 next server in the circuit. Tor as described herein, Tarzan, MorphMix,
326 <b>Cebolla</b> [<a href="#cebolla" name="CITEcebolla">9</a>], and Rennhard's <b>Anonymity Network</b> [<a href="#anonnet" name="CITEanonnet">44</a>]
327 build circuits
328 in stages, extending them one hop at a time.
330 approach enables perfect forward secrecy.
333 Circuit-based designs must choose which protocol layer
334 to anonymize. They may intercept IP packets directly, and
335 relay them whole (stripping the source address) along the
336 circuit [<a href="#freedom2-arch" name="CITEfreedom2-arch">8</a>,<a href="#tarzan:ccs02" name="CITEtarzan:ccs02">24</a>]. Like
337 Tor, they may accept TCP streams and relay the data in those streams,
338 ignoring the breakdown of that data into TCP
339 segments [<a href="#morphmix:fc04" name="CITEmorphmix:fc04">43</a>,<a href="#anonnet" name="CITEanonnet">44</a>]. Finally, like Crowds, they may accept
340 application-level protocols such as HTTP and relay the application
341 requests themselves.
342 Making this protocol-layer decision requires a compromise between flexibility
343 and anonymity. For example, a system that understands HTTP
344 can strip
345 identifying information from requests, can take advantage of caching
346 to limit the number of requests that leave the network, and can batch
347 or encode requests to minimize the number of connections.
348 On the other hand, an IP-level anonymizer can handle nearly any protocol,
349 even ones unforeseen by its designers (though these systems require
350 kernel-level modifications to some operating systems, and so are more
351 complex and less portable). TCP-level anonymity networks like Tor present
352 a middle approach: they are application neutral (so long as the
353 application supports, or can be tunneled across, TCP), but by treating
354 application connections as data streams rather than raw TCP packets,
355 they avoid the inefficiencies of tunneling TCP over
356 TCP.
359 Distributed-trust anonymizing systems need to prevent attackers from
360 adding too many servers and thus compromising user paths.
361 Tor relies on a small set of well-known directory servers, run by
362 independent parties, to decide which nodes can
363 join. Tarzan and MorphMix allow unknown users to run servers, and use
364 a limited resource (like IP addresses) to prevent an attacker from
365 controlling too much of the network. Crowds suggests requiring
366 written, notarized requests from potential crowd members.
369 Anonymous communication is essential for censorship-resistant
370 systems like Eternity [<a href="#eternity" name="CITEeternity">2</a>], Free Haven [<a href="#freehaven-berk" name="CITEfreehaven-berk">19</a>],
371 Publius [<a href="#publius" name="CITEpublius">53</a>], and Tangler [<a href="#tangler" name="CITEtangler">52</a>]. Tor's rendezvous
372 points enable connections between mutually anonymous entities; they
373 are a building block for location-hidden servers, which are needed by
374 Eternity and Free Haven.
382 </a>
386 Like other low-latency anonymity designs, Tor seeks to frustrate
387 attackers from linking communication partners, or from linking
388 multiple communications to or from a single user. Within this
389 main goal, however, several considerations have directed
390 Tor's evolution.
394 real world. Thus it
395 must not be expensive to run (for example, by requiring more bandwidth
396 than volunteers are willing to provide); must not place a heavy
397 liability burden on operators (for example, by allowing attackers to
398 implicate onion routers in illegal activities); and must not be
399 difficult or expensive to implement (for example, by requiring kernel
400 patches, or separate proxies for every protocol). We also cannot
401 require non-anonymous parties (such as websites)
402 to run our software. (Our rendezvous point design does not meet
403 this goal for non-anonymous users talking to hidden servers,
408 anonymity systems hide users among users, a system with fewer users
409 provides less anonymity. Usability is thus not only a convenience:
410 it is a security requirement [<a href="#econymics" name="CITEeconymics">1</a>,<a href="#back01" name="CITEback01">5</a>]. Tor should
411 therefore not
412 require modifying familiar applications; should not introduce prohibitive
413 delays;
414 and should require as few configuration decisions
415 as possible. Finally, Tor should be easily implementable on all common
416 platforms; we cannot require users to change their operating system
417 to be anonymous. (Tor currently runs on Win32, Linux,
418 Solaris, BSD-style Unix, MacOS X, and probably others.)
422 so Tor can serve as a test-bed for future research.
423 Many of the open problems in low-latency anonymity
424 networks, such as generating dummy traffic or preventing Sybil
425 attacks [<a href="#sybil" name="CITEsybil">22</a>], may be solvable independently from the issues
426 solved by
427 Tor. Hopefully future systems will not need to reinvent Tor's design.
431 parameters must be well-understood. Additional features impose implementation
432 and complexity costs; adding unproven techniques to the design threatens
433 deployability, readability, and ease of security analysis. Tor aims to
434 deploy a simple and stable system that integrates the best accepted
435 approaches to protecting anonymity.<br />
439 </a><br />
440 In favoring simple, deployable designs, we have explicitly deferred
441 several possible goals, either because they are solved elsewhere, or because
442 they are not yet solved.
446 decentralized peer-to-peer environments with thousands of short-lived
447 servers, many of which may be controlled by an adversary. This approach
448 is appealing, but still has many open
449 problems [<a href="#tarzan:ccs02" name="CITEtarzan:ccs02">24</a>,<a href="#morphmix:fc04" name="CITEmorphmix:fc04">43</a>].
453 to completely solve end-to-end timing or intersection
454 attacks. Some approaches, such as having users run their own onion routers,
455 may help;
460 normalization</em> like Privoxy or the Anonymizer. If senders want anonymity from
461 responders while using complex and variable
462 protocols like HTTP, Tor must be layered with a filtering proxy such
463 as Privoxy to hide differences between clients, and expunge protocol
464 features that leak identity.
465 Note that by this separation Tor can also provide services that
466 are anonymous to the network yet authenticated to the responder, like
467 SSH. Similarly, Tor does not integrate
468 tunneling for non-stream-based protocols like UDP; this must be
469 provided by an external service if appropriate.
473 to the network.
479 </a>
482 A global passive adversary is the most commonly assumed threat when
483 analyzing theoretical anonymity designs. But like all practical
484 low-latency systems, Tor does not protect against such a strong
485 adversary. Instead, we assume an adversary who can observe some fraction
486 of network traffic; who can generate, modify, delete, or delay
487 traffic; who can operate onion routers of his own; and who can
488 compromise some fraction of the onion routers.
491 In low-latency anonymity systems that use layered encryption, the
492 adversary's typical goal is to observe both the initiator and the
493 responder. By observing both ends, passive attackers can confirm a
494 suspicion that Alice is
495 talking to Bob if the timing and volume patterns of the traffic on the
496 connection are distinct enough; active attackers can induce timing
497 signatures on the traffic to force distinct patterns. Rather
499 we aim to prevent <em>traffic
500 analysis</em> attacks, where the adversary uses traffic patterns to learn
501 which points in the network he should attack.
504 Our adversary might try to link an initiator Alice with her
505 communication partners, or try to build a profile of Alice's
506 behavior. He might mount passive attacks by observing the network edges
507 and correlating traffic entering and leaving the network — by
508 relationships in packet timing, volume, or externally visible
509 user-selected
510 options. The adversary can also mount active attacks by compromising
511 routers or keys; by replaying traffic; by selectively denying service
512 to trustworthy routers to move users to
513 compromised routers, or denying service to users to see if traffic
514 elsewhere in the
515 network stops; or by introducing patterns into traffic that can later be
516 detected. The adversary might subvert the directory servers to give users
517 differing views of network state. Additionally, he can try to decrease
518 the network's reliability by attacking nodes or by performing antisocial
519 activities from reliable nodes and trying to get them taken down — making
520 the network unreliable flushes users to other less anonymous
521 systems, where they may be easier to attack. We summarize
523 each of these attacks.
531 </a>
534 The Tor network is an overlay network; each onion router (OR)
535 runs as a normal
536 user-level process without any special privileges.
538 connection to every other onion router.
539 Each user
540 runs local software called an onion proxy (OP) to fetch directories,
541 establish circuits across the network,
542 and handle connections from user applications. These onion proxies accept
543 TCP streams and multiplex them across the circuits. The onion
544 router on the other side
545 of the circuit connects to the requested destinations
546 and relays data.
549 Each onion router maintains a long-term identity key and a short-term
550 onion key. The identity
551 key is used to sign TLS certificates, to sign the OR's <em>router
552 descriptor</em> (a summary of its keys, address, bandwidth, exit policy,
553 and so on), and (by directory servers) to sign directories. The onion key is used to decrypt requests
554 from users to set up a circuit and negotiate ephemeral keys.
555 The TLS protocol also establishes a short-term link key when communicating
556 between ORs. Short-term keys are rotated periodically and
557 independently, to limit the impact of key compromise.
564 describes how TCP streams are routed through the network. We address
567 Finally,
569 fairness issues.
575 </a>
578 Onion routers communicate with one another, and with users' OPs, via
579 TLS connections with ephemeral keys. Using TLS conceals the data on
580 the connection with perfect forward secrecy, and prevents an attacker
581 from modifying data on the wire or impersonating an OR.
584 Traffic passes along these connections in fixed-size cells. Each cell
585 is 512 bytes, and consists of a header and a payload. The header includes a circuit
586 identifier (circID) that specifies which circuit the cell refers to
587 (many circuits can be multiplexed over the single TLS connection), and
588 a command to describe what to do with the cell's payload. (Circuit
589 identifiers are connection-specific: each circuit has a different
590 circID on each OP/OR or OR/OR connection it traverses.)
593 which carry end-to-end stream data. The control cell commands are:
599 Relay cells have an additional header (the relay header) at the front
600 of the payload, containing a streamID (stream identifier: many streams can
601 be multiplexed over a circuit); an end-to-end checksum for integrity
602 checking; the length of the relay payload; and a relay command.
603 The entire contents of the relay header and the relay cell payload
604 are encrypted or decrypted together as the relay cell moves along the
605 circuit, using the 128-bit AES cipher in counter mode to generate a
606 cipher stream. The relay commands are: <em>relay
610 (to notify the OP that a relay begin has succeeded), <em>relay
613 (to tear down only part of the circuit, and to acknowledge), <em>relay
615 implement long-range dummies).
616 We give a visual overview of cell structure plus the details of relay
617 cell structure, and then describe each of these cell types and commands
618 in more detail below.
627 </center>
632 </a>
635 Onion Routing originally built one circuit for each
636 TCP stream. Because building a circuit can take several tenths of a
637 second (due to public-key cryptography and network latency),
638 this design imposed high costs on applications like web browsing that
639 open many TCP streams.
642 In Tor, each circuit can be shared by many TCP streams. To avoid
643 delays, users construct circuits preemptively. To limit linkability
644 among their streams, users' OPs build a new circuit
645 periodically if the previous ones have been used,
646 and expire old used circuits that no longer have any open streams.
647 OPs consider rotating to a new circuit once a minute: thus
648 even heavy users spend negligible time
649 building circuits, but a limited number of requests can be linked
650 to each other through a given exit node. Also, because circuits are built
651 in the background, OPs can recover from failed circuit creation
652 without harming user experience.<br />
662 </a>
663 </center>
667 <br />
668 A user's OP constructs circuits incrementally, negotiating a
669 symmetric key with each OR on the circuit, one hop at a time. To begin
670 creating a new circuit, the OP (call her Alice) sends a
672 (She chooses a new
675 payload contains the first half of the Diffie-Hellman handshake
681 Once the circuit has been established, Alice and Bob can send one
682 another relay cells encrypted with the negotiated
684 the next section.
688 to Bob, specifying the address of the next OR (call her Carol), and
692 between him and Carol. Alice never needs to know this circID; only Bob
694 his connection with Carol.)
697 the circuit is extended to Carol, and Alice and Carol share a common key
701 To extend the circuit to a third node or beyond, Alice
702 proceeds as above, always telling the last node in the circuit to
703 extend one hop further.
706 This circuit-level handshake protocol achieves unilateral entity
707 authentication (Alice knows she's handshaking with the OR, but
708 the OR doesn't care who is opening the circuit — Alice uses no public key
709 and remains anonymous) and unilateral key authentication
710 (Alice and the OR agree on a key, and Alice knows only the OR learns
711 it). It also achieves forward
712 secrecy and key freshness. More formally, the protocol is as follows
715 > is concatenation):
719 </a>
720 <table>
721 <tr><td align="right">Alice </td><td align="center">-> </td><td align="center">Bob </td><td>: E<sub>PK<sub>Bob</sub></sub>(g<sup>x</sup>) </td></tr>
722 <tr><td align="right">Bob </td><td align="center">-> </td><td align="center">Alice </td><td>: g<sup>y</sup>, H(K <font face="symbol">|</font
724 </td></tr></table>
729 and who chose y. We use PK encryption in the first step
730 (rather than, say, using the first two steps of STS, which has a
731 signature in the second step) because a single cell is too small to
732 hold both a public key and a signature. Preliminary analysis with the
733 NRL protocol analyzer [<a href="#meadows96" name="CITEmeadows96">35</a>] shows this protocol to be
734 secure (including perfect forward secrecy) under the
735 traditional Dolev-Yao model.<br />
739 Once Alice has established the circuit (so she shares keys with each
740 OR on the circuit), she can send relay cells.
741 Upon receiving a relay
742 cell, an OR looks up the corresponding circuit, and decrypts the relay
743 header and payload with the session key for that circuit.
744 If the cell is headed away from Alice the OR then checks whether the
745 decrypted cell has a valid digest (as an optimization, the first
746 two bytes of the integrity check are zero, so in most cases we can avoid
747 computing the hash).
748 If valid, it accepts the relay cell and processes it as described
749 below. Otherwise,
750 the OR looks up the circID and OR for the
751 next step in the circuit, replaces the circID as appropriate, and
752 sends the decrypted relay cell to the next OR. (If the OR at the end
753 of the circuit receives an unrecognized relay cell, an error has
754 occurred, and the circuit is torn down.)
757 OPs treat incoming relay cells similarly: they iteratively unwrap the
758 relay header and payload with the session keys shared with each
759 OR on the circuit, from the closest to farthest.
760 If at any stage the digest is valid, the cell must have
761 originated at the OR whose encryption has just been removed.
764 To construct a relay cell addressed to a given OR, Alice assigns the
765 digest, and then iteratively
766 encrypts the cell payload (that is, the relay header and payload) with
767 the symmetric key of each hop up to that OR. Because the digest is
768 encrypted to a different value at each step, only at the targeted OR
771 allows Alice's streams to exit at different ORs on a single circuit.
772 Alice may choose different exit points because of their exit policies,
773 or to keep the ORs from knowing that two streams
774 originate from the same person.
777 When an OR later replies to Alice with a relay cell, it
778 encrypts the cell's relay header and payload with the single key it
779 shares with Alice, and sends the cell back toward Alice along the
780 circuit. Subsequent ORs add further layers of encryption as they
781 relay the cell back to Alice.
787 forward. But just as circuits are built incrementally, they can also
788 be torn down incrementally: Alice can send a <em>relay
789 truncate</em> cell to a single OR on a circuit. That OR then sends a
792 different nodes, without signaling to the intermediate nodes (or
793 a limited observer) that she has changed her circuit.
794 Similarly, if a node on the circuit goes down, the adjacent
796 "break a node and see which circuits go down"
803 </a>
806 When Alice's application wants a TCP connection to a given
807 address and port, it asks the OP (via SOCKS) to make the
808 connection. The OP chooses the newest open circuit (or creates one if
809 needed), and chooses a suitable OR on that circuit to be the
810 exit node (usually the last node, but maybe others due to exit policy
813 using a new random streamID. Once the
814 exit node connects to the remote host, it responds
816 SOCKS reply to notify the application of its success. The OP
817 now accepts data from the application's TCP stream, packaging it into
819 the chosen OR.
822 There's a catch to using SOCKS, however — some applications pass the
823 alphanumeric hostname to the Tor client, while others resolve it into
824 an IP address first and then pass the IP address to the Tor client. If
825 the application does DNS resolution first, Alice thereby reveals her
826 destination to the remote DNS server, rather than sending the hostname
827 through the Tor network to be resolved at the far end. Common applications
828 like Mozilla and SSH have this flaw.
831 With Mozilla, the flaw is easy to address: the filtering HTTP
832 proxy called Privoxy gives a hostname to the Tor client, so Alice's
833 computer never does DNS resolution.
834 But a portable general solution, such as is needed for
835 SSH, is
836 an open problem. Modifying or replacing the local nameserver
837 can be invasive, brittle, and unportable. Forcing the resolver
838 library to prefer TCP rather than UDP is hard, and also has
839 portability problems. Dynamically intercepting system calls to the
840 resolver library seems a promising direction. We could also provide
842 Tor network. Currently, we encourage the use of privacy-aware proxies
843 like Privoxy wherever possible.
846 Closing a Tor stream is analogous to closing a TCP stream: it uses a
847 two-step handshake for normal operation, or a one-step handshake for
848 errors. If the stream closes abnormally, the adjacent node simply sends a
852 all relay cells use layered encryption, only the destination OR knows
853 that a given relay cell is a request to close a stream. This two-step
854 handshake allows Tor to support TCP-based applications that use half-closed
855 connections.
861 </a>
864 Because the old Onion Routing design used a stream cipher without integrity
865 checking, traffic was
866 vulnerable to a malleability attack: though the attacker could not
867 decrypt cells, any changes to encrypted data
868 would create corresponding changes to the data leaving the network.
869 This weakness allowed an adversary who could guess the encrypted content
870 to change a padding cell to a destroy
872 adversary's webserver; or change an FTP command from
874 adversary could do this, because the link encryption similarly used a
875 stream cipher.)
878 Because Tor uses TLS on its links, external adversaries cannot modify
879 data. Addressing the insider malleability attack, however, is
880 more complex.
883 We could do integrity checking of the relay cells at each hop, either
884 by including hashes or by using an authenticating cipher mode like
885 EAX [<a href="#eax" name="CITEeax">6</a>], but there are some problems. First, these approaches
886 impose a message-expansion overhead at each hop, and so we would have to
887 either leak the path length or waste bytes by padding to a maximum
888 path length. Second, these solutions can only verify traffic coming
889 from Alice: ORs would not be able to produce suitable hashes for
890 the intermediate hops, since the ORs on a circuit do not know the
891 other ORs' session keys. Third, we have already accepted that our design
892 is vulnerable to end-to-end timing attacks; so tagging attacks performed
893 within the circuit provide no additional information to the attacker.
896 Thus, we check integrity only at the edges of each stream. (Remember that
897 in our leaky-pipe circuit topology, a stream's edge could be any hop
898 in the circuit.) When Alice
899 negotiates a key with a new hop, they each initialize a SHA-1
900 digest with a derivative of that key,
901 thus beginning with randomness that only the two of them know.
902 Then they each incrementally add to the SHA-1 digest the contents of
903 all relay cells they create, and include with each relay cell the
904 first four bytes of the current digest. Each also keeps a SHA-1
905 digest of data received, to verify that the received hashes are correct.
908 To be sure of removing or modifying a cell, the attacker must be able
909 to deduce the current digest state (which depends on all
910 traffic between Alice and Bob, starting with their negotiated key).
911 Attacks on SHA-1 where the adversary can incrementally add to a hash
912 to produce a new valid hash don't work, because all hashes are
913 end-to-end encrypted across the circuit. The computational overhead
914 of computing the digests is minimal compared to doing the AES
915 encryption performed at each hop of the circuit. We use only four
916 bytes per cell to minimize overhead; the chance that an adversary will
917 correctly guess a valid hash
918 is
919 acceptably low, given that the OP or OR tear down the circuit if they
920 receive a bad hash.
926 </a>
929 Volunteers are more willing to run services that can limit
930 their bandwidth usage. To accommodate them, Tor servers use a
932 enforce a long-term average rate of incoming bytes, while still
933 permitting short-term bursts above the allowed bandwidth.
938 Because the Tor protocol outputs about the same number of bytes as it
939 takes in, it is sufficient in practice to limit only incoming bytes.
940 With TCP streams, however, the correspondence is not one-to-one:
941 relaying a single incoming byte can require an entire 512-byte cell.
942 (We can't just wait for more bytes, because the local application may
943 be awaiting a reply.) Therefore, we treat this case as if the entire
944 cell size had been read, regardless of the cell's fullness.
947 Further, inspired by Rennhard et al's design in [<a href="#anonnet" name="CITEanonnet">44</a>], a
948 circuit's edges can heuristically distinguish interactive streams from bulk
949 streams by comparing the frequency with which they supply cells. We can
950 provide good latency for interactive streams by giving them preferential
951 service, while still giving good overall throughput to the bulk
952 streams. Such preferential treatment presents a possible end-to-end
953 attack, but an adversary observing both
954 ends of the stream can already learn this information through timing
955 attacks.
961 </a>
964 Even with bandwidth rate limiting, we still need to worry about
965 congestion, either accidental or intentional. If enough users choose the
966 same OR-to-OR connection for their circuits, that connection can become
967 saturated. For example, an attacker could send a large file
968 through the Tor network to a webserver he runs, and then
969 refuse to read any of the bytes at the webserver end of the
970 circuit. Without some congestion control mechanism, these bottlenecks
971 can propagate back through the entire network. We don't need to
972 reimplement full TCP windows (with sequence numbers,
973 the ability to drop cells when we're full and retransmit later, and so
974 on),
975 because TCP already guarantees in-order delivery of each
976 cell.
977 We describe our response below.
981 To control a circuit's bandwidth usage, each OR keeps track of two
983 allowed to package (from incoming TCP streams) for transmission back to the OP,
985 to deliver to TCP streams outside the network. Each window is initialized
986 (say, to 1000 data cells). When a data cell is packaged or delivered,
987 the appropriate window is decremented. When an OR has received enough
990 streamID zero, it increments its packaging window. Either of these cells
991 increments the corresponding window by 100. If the packaging window
992 reaches 0, the OR stops reading from TCP connections for all streams
993 on the corresponding circuit, and sends no more relay data cells until
997 The OP behaves identically, except that it must track a packaging window
998 and a delivery window for every OR in the circuit. If a packaging window
999 reaches 0, it stops reading from streams destined for that OR.
1003 The stream-level congestion control mechanism is similar to the
1005 to implement end-to-end flow control for individual streams across
1006 circuits. Each stream begins with a packaging window (currently 500 cells),
1009 as enough cells have arrived, the stream-level congestion control also
1010 has to check whether data has been successfully flushed onto the TCP
1012 to be flushed is under some threshold (currently 10 cells' worth).
1017 These arbitrarily chosen parameters seem to give tolerable throughput
1024 </a>
1027 Rendezvous points are a building block for <em>location-hidden
1029 network. Location-hidden services allow Bob to offer a TCP
1030 service, such as a webserver, without revealing his IP address.
1031 This type of anonymity protects against distributed DoS attacks:
1032 attackers are forced to attack the onion routing network
1033 because they do not know Bob's IP address.
1036 Our design for location-hidden servers has the following goals.
1038 so an attacker cannot flood Bob simply by making many connections to him.
1040 identity even in the presence of router failure. Bob's service must
1041 not be tied to a single OR, and Bob must be able to migrate his service
1043 A social attacker
1044 should not be able to "frame" a rendezvous router by
1045 offering an illegal or disreputable location-hidden service and
1046 making observers believe the router created that service.
1048 to run special software to access location-hidden servers, we must not
1049 require them to modify their applications.
1052 We provide location-hiding for Bob by allowing him to advertise
1054 points. He may do this on any robust efficient
1055 key-value lookup system with authenticated updates, such as a
1056 distributed hash table (DHT) like CFS [<a href="#cfs:sosp01" name="CITEcfs:sosp01">11</a>].<a href="#tthFtNtAAD" name="tthFrefAAD"><sup>3</sup></a> Alice, the client, chooses an OR as her
1058 points, informs him of her rendezvous point, and then waits for him
1059 to connect to the rendezvous point. This extra level of indirection
1060 helps Bob's introduction points avoid problems associated with serving
1061 unpopular files directly (for example, if Bob serves
1062 material that the introduction point's community finds objectionable,
1063 or if Bob's service tends to get attacked by network vandals).
1064 The extra level of indirection also allows Bob to respond to some requests
1065 and ignore others.
1072 The following steps are
1073 performed on behalf of Alice and Bob by their local OPs;
1074 application integration is described more fully below.
1080 <dt><b></b></dt>
1082 <dt><b></b></dt>
1083 <dd><li>Bob chooses some introduction points, and advertises them on
1084 the lookup service, signing the advertisement with his public key. He
1085 can add more later.</dd>
1086 <dt><b></b></dt>
1087 <dd><li>Bob builds a circuit to each of his introduction points, and tells
1088 them to wait for requests.</dd>
1089 <dt><b></b></dt>
1090 <dd><li>Alice learns about Bob's service out of band (perhaps Bob told her,
1091 or she found it on a website). She retrieves the details of Bob's
1092 service from the lookup service. If Alice wants to access Bob's
1093 service anonymously, she must connect to the lookup service via Tor.</dd>
1094 <dt><b></b></dt>
1095 <dd><li>Alice chooses an OR as the rendezvous point (RP) for her connection to
1096 Bob's service. She builds a circuit to the RP, and gives it a
1098 <dt><b></b></dt>
1099 <dd><li>Alice opens an anonymous stream to one of Bob's introduction
1100 points, and gives it a message (encrypted with Bob's public key)
1101 telling it about herself,
1102 her RP and rendezvous cookie, and the
1103 start of a DH
1104 handshake. The introduction point sends the message to Bob.</dd>
1105 <dt><b></b></dt>
1106 <dd><li>If Bob wants to talk to Alice, he builds a circuit to Alice's
1107 RP and sends the rendezvous cookie, the second half of the DH
1108 handshake, and a hash of the session
1109 key they now share. By the same argument as in
1111 shares the key only with Bob.</dd>
1112 <dt><b></b></dt>
1113 <dd><li>The RP connects Alice's circuit to Bob's. Note that RP can't
1114 recognize Alice, Bob, or the data they transmit.</dd>
1115 <dt><b></b></dt>
1117 arrives at Bob's OP, which connects to Bob's
1118 webserver.</dd>
1119 <dt><b></b></dt>
1120 <dd><li>An anonymous stream has been established, and Alice and Bob
1121 communicate as normal.
1122 </dd>
1123 </dl>
1126 When establishing an introduction point, Bob provides the onion router
1127 with the public key identifying his service. Bob signs his
1128 messages, so others cannot usurp his introduction point
1129 in the future. He uses the same public key to establish the other
1130 introduction points for his service, and periodically refreshes his
1131 entry in the lookup service.
1134 The message that Alice gives
1135 the introduction point includes a hash of Bob's public key and an optional initial authorization token (the
1136 introduction point can do prescreening, for example to block replays). Her
1137 message to Bob may include an end-to-end authorization token so Bob
1138 can choose whether to respond.
1139 The authorization tokens can be used to provide selective access:
1140 important users can get uninterrupted access.
1141 During normal situations, Bob's service might simply be offered
1142 directly from mirrors, while Bob gives out tokens to high-priority users. If
1143 the mirrors are knocked down,
1144 those users can switch to accessing Bob's service via
1145 the Tor rendezvous system.
1148 Bob's introduction points are themselves subject to DoS — he must
1149 open many introduction points or risk such an attack.
1150 He can provide selected users with a current list or future schedule of
1151 unadvertised introduction points;
1152 this is most practical
1153 if there is a stable and large group of introduction points
1154 available. Bob could also give secret public keys
1155 for consulting the lookup service. All of these approaches
1156 limit exposure even when
1157 some selected users collude in the DoS.
1164 Bob configures his onion proxy to know the local IP address and port of his
1165 service, a strategy for authorizing clients, and his public key. The onion
1166 proxy anonymously publishes a signed statement of Bob's
1167 public key, an expiration time, and
1168 the current introduction points for his service onto the lookup service,
1169 indexed
1170 by the hash of his public key. Bob's webserver is unmodified,
1171 and doesn't even know that it's hidden behind the Tor network.
1174 Alice's applications also work unchanged — her client interface
1175 remains a SOCKS proxy. We encode all of the necessary information
1176 into the fully qualified domain name (FQDN) Alice uses when establishing her
1177 connection. Location-hidden services use a virtual top level domain
1180 the public key. Alice's onion proxy
1181 examines addresses; if they're destined for a hidden server, it decodes
1182 the key and starts the rendezvous as described above.
1189 Rendezvous points in low-latency anonymity systems were first
1190 described for use in ISDN telephony [<a href="#jerichow-jsac98" name="CITEjerichow-jsac98">30</a>,<a href="#isdn-mixes" name="CITEisdn-mixes">38</a>].
1191 Later low-latency designs used rendezvous points for hiding location
1192 of mobile phones and low-power location
1193 trackers [<a href="#federrath-ih96" name="CITEfederrath-ih96">23</a>,<a href="#reed-protocols97" name="CITEreed-protocols97">40</a>]. Rendezvous for
1194 anonymizing low-latency
1195 Internet connections was suggested in early Onion Routing
1196 work [<a href="#or-ih96" name="CITEor-ih96">27</a>], but the first published design was by Ian
1198 ours in three ways. First, Goldberg suggests that Alice should manually
1199 hunt down a current location of the service via Gnutella; our approach
1200 makes lookup transparent to the user, as well as faster and more robust.
1201 Second, in Tor the client and server negotiate session keys
1202 with Diffie-Hellman, so plaintext is not exposed even at the rendezvous
1203 point. Third,
1204 our design minimizes the exposure from running the
1205 service, to encourage volunteers to offer introduction and rendezvous
1206 services. Tor's introduction points do not output any bytes to the
1207 clients; the rendezvous points don't know the client or the server,
1208 and can't read the data being transmitted. The indirection scheme is
1209 also designed to include authentication/authorization — if Alice doesn't
1210 include the right cookie with her request for service, Bob need not even
1211 acknowledge his existence.
1217 </a>
1223 </a>
1226 Providing Tor as a public service creates many opportunities for
1227 denial-of-service attacks against the network. While
1228 flow control and rate limiting (discussed in
1230 bandwidth than routers are willing to provide, opportunities remain for
1231 users to
1232 consume more network resources than their fair share, or to render the
1233 network unusable for others.
1236 First of all, there are several CPU-consuming denial-of-service
1237 attacks wherein an attacker can force an OR to perform expensive
1238 cryptographic operations. For example, an attacker can
1239 fake the start of a TLS handshake, forcing the OR to carry out its
1240 (comparatively expensive) half of the handshake at no real computational
1241 cost to the attacker.
1244 We have not yet implemented any defenses for these attacks, but several
1245 approaches are possible. First, ORs can
1246 require clients to solve a puzzle [<a href="#puzzles-tls" name="CITEpuzzles-tls">16</a>] while beginning new
1248 tokens are easy to verify and computationally expensive to produce, this
1249 approach limits the attack multiplier. Additionally, ORs can limit
1251 so that
1252 the computational work of processing them does not drown out the
1253 symmetric cryptography operations that keep cells
1254 flowing. This rate limiting could, however, allow an attacker
1255 to slow down other users when they build new circuits.
1260 Adversaries can also attack the Tor network's hosts and network
1261 links. Disrupting a single circuit or link breaks all streams passing
1262 along that part of the circuit. Users similarly lose service
1263 when a router crashes or its operator restarts it. The current
1264 Tor design treats such attacks as intermittent network failures, and
1265 depends on users and applications to respond or recover as appropriate. A
1266 future design could use an end-to-end TCP-like acknowledgment protocol,
1267 so no streams are lost unless the entry or exit point is
1268 disrupted. This solution would require more buffering at the network
1269 edges, however, and the performance and anonymity implications from this
1270 extra complexity still require investigation.
1276 </a>
1281 Exit abuse is a serious barrier to wide-scale Tor deployment. Anonymity
1282 presents would-be vandals and abusers with an opportunity to hide
1283 the origins of their activities. Attackers can harm the Tor network by
1284 implicating exit servers for their abuse. Also, applications that commonly
1285 use IP-based authentication (such as institutional mail or webservers)
1286 can be fooled by the fact that anonymous connections appear to originate
1287 at the exit OR.
1290 We stress that Tor does not enable any new class of abuse. Spammers
1291 and other attackers already have access to thousands of misconfigured
1292 systems worldwide, and the Tor network is far from the easiest way
1293 to launch attacks.
1294 But because the
1295 onion routers can be mistaken for the originators of the abuse,
1296 and the volunteers who run them may not want to deal with the hassle of
1297 explaining anonymity networks to irate administrators, we must block or limit
1298 abuse through the Tor network.
1302 describes to which external addresses and ports the router will
1306 nodes that only connect to a local host or network. A private
1307 exit can allow a client to connect to a given host or
1308 network more securely — an external adversary cannot eavesdrop traffic
1309 between the private exit and the final destination, and so is less sure of
1310 Alice's destination and activities. Most onion routers in the current
1311 network function as
1313 but prevent access to certain abuse-prone addresses and services such
1314 as SMTP.
1315 The OR might also be able to authenticate clients to
1316 prevent exit abuse without harming anonymity [<a href="#or-discex00" name="CITEor-discex00">48</a>].
1321 Many administrators use port restrictions to support only a
1322 limited set of services, such as HTTP, SSH, or AIM.
1323 This is not a complete solution, of course, since abuse opportunities for these
1324 protocols are still well known.
1327 We have not yet encountered any abuse in the deployed network, but if
1328 we do we should consider using proxies to clean traffic for certain
1329 protocols as it leaves the network. For example, much abusive HTTP
1330 behavior (such as exploiting buffer overflows or well-known script
1331 vulnerabilities) can be detected in a straightforward manner.
1332 Similarly, one could run automatic spam filtering software (such as
1333 SpamAssassin) on email exiting the OR network.
1336 ORs may also rewrite exiting traffic to append
1337 headers or other information indicating that the traffic has passed
1338 through an anonymity service. This approach is commonly used
1339 by email-only anonymity systems. ORs can also
1341 alert abuse targets to the nature of the anonymous traffic.
1344 A mixture of open and restricted exit nodes allows the most
1345 flexibility for volunteers running servers. But while having many
1346 middleman nodes provides a large and robust network,
1347 having only a few exit nodes reduces the number of points
1348 an adversary needs to monitor for traffic analysis, and places a
1349 greater burden on the exit nodes. This tension can be seen in the
1350 Java Anon Proxy
1351 cascade model, wherein only one node in each cascade needs to handle
1352 abuse complaints — but an adversary only needs to observe the entry
1353 and exit of a cascade to perform traffic analysis on all that
1354 cascade's users. The hydra model (many entries, few exits) presents a
1355 different compromise: only a few exit nodes are needed, but an
1356 adversary needs to work harder to watch all the clients; see
1360 Finally, we note that exit abuse must not be dismissed as a peripheral
1361 issue: when a system's public image suffers, it can reduce the number
1362 and diversity of that system's users, and thereby reduce the anonymity
1363 of the system itself. Like usability, public perception is a
1364 security parameter. Sadly, preventing abuse of open exit nodes is an
1365 unsolved problem, and will probably remain an arms race for the
1366 foreseeable future. The abuse problems faced by Princeton's CoDeeN
1367 project [<a href="#darkside" name="CITEdarkside">37</a>] give us a glimpse of likely issues.
1373 </a>
1376 First-generation Onion Routing designs [<a href="#freedom2-arch" name="CITEfreedom2-arch">8</a>,<a href="#or-jsac98" name="CITEor-jsac98">41</a>] used
1377 in-band network status updates: each router flooded a signed statement
1378 to its neighbors, which propagated it onward. But anonymizing networks
1379 have different security goals than typical link-state routing protocols.
1380 For example, delays (accidental or intentional)
1381 that can cause different parts of the network to have different views
1382 of link-state and topology are not only inconvenient: they give
1383 attackers an opportunity to exploit differences in client knowledge.
1384 We also worry about attacks to deceive a
1385 client about the router membership list, topology, or current network
1387 adversary to efficiently deploy resources
1391 Tor uses a small group of redundant, well-known onion routers to
1392 track changes in network topology and node state, including keys and
1394 server, so clients can fetch current network state
1395 and router lists, and so other ORs can upload
1396 state information. Onion routers periodically publish signed
1397 statements of their state to each directory server. The directory servers
1398 combine this information with their own views of network liveness,
1400 network state. Client software is
1401 pre-loaded with a list of the directory servers and their keys,
1402 to bootstrap each client's view of the network.
1405 When a directory server receives a signed statement for an OR, it
1406 checks whether the OR's identity key is recognized. Directory
1407 servers do not advertise unrecognized ORs — if they did,
1408 an adversary could take over the network by creating many
1409 servers [<a href="#sybil" name="CITEsybil">22</a>]. Instead, new nodes must be approved by the
1410 directory
1411 server administrator before they are included. Mechanisms for automated
1412 node approval are an area of active research, and are discussed more
1416 Of course, a variety of attacks remain. An adversary who controls
1417 a directory server can track clients by providing them different
1418 information — perhaps by listing only nodes under its control, or by
1419 informing only certain clients about a given node. Even an external
1420 adversary can exploit differences in client knowledge: clients who use
1421 a node listed on one directory server but not the others are vulnerable.
1424 Thus these directory servers must be synchronized and redundant, so
1425 that they can agree on a common directory. Clients should only trust
1426 this directory if it is signed by a threshold of the directory
1427 servers.
1430 The directory servers in Tor are modeled after those in
1431 Mixminion [<a href="#minion-design" name="CITEminion-design">15</a>], but our situation is easier. First,
1432 we make the
1433 simplifying assumption that all participants agree on the set of
1434 directory servers. Second, while Mixminion needs to predict node
1435 behavior, Tor only needs a threshold consensus of the current
1436 state of the network. Third, we assume that we can fall back to the
1437 human administrators to discover and resolve problems when a consensus
1438 directory cannot be reached. Since there are relatively few directory
1440 we can afford operations like broadcast to simplify the consensus-building
1441 protocol.
1444 To avoid attacks where a router connects to all the directory servers
1445 but refuses to relay traffic from other routers, the directory servers
1446 must also build circuits and use them to anonymously test router
1447 reliability [<a href="#mix-acc" name="CITEmix-acc">18</a>]. Unfortunately, this defense is not yet
1448 designed or
1449 implemented.
1452 Using directory servers is simpler and more flexible than flooding.
1453 Flooding is expensive, and complicates the analysis when we
1454 start experimenting with non-clique network topologies. Signed
1455 directories can be cached by other
1456 onion routers,
1457 so directory servers are not a performance
1458 bottleneck when we have many users, and do not aid traffic analysis by
1459 forcing clients to announce their existence to any
1460 central point.
1466 </a>
1469 Below we summarize a variety of attacks, and discuss how well our
1470 design withstands them.<br />
1475 will not reveal her destination or data, but it will
1476 reveal traffic patterns (both sent and received). Profiling via user
1477 connection patterns requires further processing, because multiple
1478 application streams may be operating simultaneously or in series over
1479 a single circuit.
1483 connections to responders may not be (indeed, the responding website
1484 itself may be hostile). While filtering content is not a primary goal
1485 of Onion Routing, Tor can directly use Privoxy and related
1486 filtering services to anonymize application data streams.
1490 configuration options. For example, clients concerned about request
1491 linkability should rotate circuits more often than those concerned
1492 about traceability. Allowing choice may attract users with different
1493 needs; but clients who are
1494 in the minority may lose more anonymity by appearing distinct than they
1499 such correlations. An attacker watching patterns of
1500 traffic at the initiator and the responder will be
1501 able to confirm the correspondence with high probability. The
1502 greatest protection currently available against such confirmation is to hide
1503 the connection between the onion proxy and the first Tor node,
1504 by running the OP on the Tor node or behind a firewall. This approach
1505 requires an observer to separate traffic originating at the onion
1506 router from traffic passing through it: a global observer can do this,
1507 but it might be beyond a limited observer's capabilities.
1511 will also be effective in confirming
1512 endpoints of a stream. However, even without padding, we may have some
1513 limited protection: the leaky pipe topology means different numbers
1514 of packets may enter one end of a circuit than exit at the other.
1518 attacks above are traffic confirmation attacks,
1519 which puts them outside our design goals. There is also
1520 a passive traffic analysis attack that is potentially effective.
1521 Rather than searching exit connections for timing and volume
1522 correlations, the adversary may build up a database of
1523 "fingerprints" containing file sizes and access patterns for
1524 targeted websites. He can later confirm a user's connection to a given
1525 site simply by consulting the database. This attack has
1526 been shown to be effective against SafeWeb [<a href="#hintz-pet02" name="CITEhintz-pet02">29</a>].
1527 It may be less effective against Tor, since
1528 streams are multiplexed within the same circuit, and
1529 fingerprinting will be limited to
1530 the granularity of cells (currently 512 bytes). Additional
1531 defenses could include
1532 larger cell sizes, padding schemes to group websites
1533 into large sets, and link
1539 see control cells and encrypted relay cells on every circuit on that
1540 connection; learning a circuit
1541 session key lets him unwrap one layer of the encryption. An attacker
1542 who learns an OR's TLS private key can impersonate that OR for the TLS
1543 key's lifetime, but he must
1545 perfect forward secrecy, he cannot hijack already established circuits
1546 without also compromising their session keys). Periodic key rotation
1547 limits the window of opportunity for these attacks. On the other hand,
1548 an attacker who learns a node's identity key can replace that node
1549 indefinitely by sending new forged descriptors to the directory servers.
1553 compromise ORs (by system intrusion, legal coercion, or extralegal
1554 coercion) could march down the circuit compromising the
1555 nodes until he reaches the end. Unless the adversary can complete
1556 this attack within the lifetime of the circuit, however, the ORs
1557 will have discarded the necessary information before the attack can
1558 be completed. (Thanks to the perfect forward secrecy of session
1559 keys, the attacker cannot force nodes to decrypt recorded
1560 traffic once the circuits have been closed.) Additionally, building
1561 circuits that cross jurisdictions can make legal coercion
1563 arbitrage." The Java Anon Proxy project recently experienced the
1564 need for this approach, when
1565 a German court forced them to add a backdoor to
1570 trivially learns the timing patterns of users connecting to it, and
1571 can introduce arbitrary patterns in its responses.
1572 End-to-end attacks become easier: if the adversary can induce
1573 users to connect to his webserver (perhaps by advertising
1574 content targeted to those users), he now holds one end of their
1575 connection. There is also a danger that application
1576 protocols and associated programs can be induced to reveal information
1577 about the initiator. Tor depends on Privoxy and similar protocol cleaners
1578 to solve this latter problem.
1582 nearly always run their own local onion proxy. However, in some
1583 settings, it may be necessary for the proxy to run
1584 remotely — typically, in institutions that want
1585 to monitor the activity of those connecting to the proxy.
1586 Compromising an onion proxy compromises all future connections
1587 through it.
1591 of the Tor network can increase the value of this traffic
1592 by attacking non-observed nodes to shut them down, reduce
1593 their reliability, or persuade users that they are not trustworthy.
1594 The best defense here is robustness.
1598 an isolated hostile node can create circuits through itself, or alter
1599 traffic patterns to affect traffic at other nodes. Nonetheless, a hostile
1600 node must be immediately adjacent to both endpoints to compromise the
1601 anonymity of a circuit. If an adversary can
1602 run multiple ORs, and can persuade the directory servers
1603 that those ORs are trustworthy and independent, then occasionally
1604 some user will choose one of those ORs for the start and another
1605 as the end of a circuit. If an adversary
1608 adversary
1609 could still attract a disproportionately large amount of traffic
1610 by running an OR with a permissive exit policy, or by
1611 degrading the reliability of other routers.
1615 version of passive timing attacks already discussed earlier.
1619 cell by altering it. If the
1620 stream were, for example, an unencrypted request to a Web site,
1621 the garbled content coming out at the appropriate time would confirm
1622 the association. However, integrity checks on cells prevent
1623 this attack.
1627 relaying an unauthenticated protocol like HTTP, a hostile exit node
1628 can impersonate the target server. Clients
1629 should prefer protocols with end-to-end authentication.
1633 to replay attacks. Tor is not; replaying one side of a handshake
1634 will result in a different negotiated session key, and so the rest
1635 of the recorded session can't be used.
1639 socially disapproved acts, to bring the
1640 network into disrepute and get its operators to shut it down.
1641 Exit policies reduce the possibilities for abuse, but
1642 ultimately the network requires volunteers who can tolerate
1643 some political heat.
1647 into running subverted Tor software that did not, in fact, anonymize
1648 their connections — or worse, could trick ORs into running weakened
1649 software that provided users with less anonymity. We address this
1650 problem (but do not solve it completely) by signing all Tor releases
1651 with an official public key, and including an entry in the directory
1652 that lists which versions are currently believed to be secure. To
1653 prevent an attacker from subverting the official release itself
1654 (through threats, bribery, or insider attacks), we provide all
1655 releases in source code form, encourage source audits, and
1656 frequently warn our users never to trust any software (even from
1657 us) that comes without source.<br />
1662 servers disappear, the others still decide on a valid
1663 directory. So long as any directory servers remain in operation,
1664 they will still broadcast their views of the network and generate a
1665 consensus directory. (If more than half are destroyed, this
1666 directory will not, however, have enough signatures for clients to
1667 use it automatically; human intervention will be necessary for
1668 clients to decide whether to trust the resulting directory.)
1672 an attacker can partially influence the final directory. Since ORs
1673 are included or excluded by majority vote, the corrupt directory can
1674 at worst cast a tie-breaking vote to decide whether to include
1675 marginal ORs. It remains to be seen how often such marginal cases
1676 occur in practice.
1680 more than half the directory servers can include as many compromised
1681 ORs in the final directory as he wishes. We must ensure that directory
1682 server operators are independent and attack-resistant.
1686 agreement protocol assumes that directory server operators agree on
1687 the set of directory servers. An adversary who can persuade some
1688 of the directory server operators to distrust one another could
1689 split the quorum into mutually hostile camps, thus partitioning
1690 users based on which directory they use. Tor does not address
1691 this attack.
1695 Our threat model explicitly assumes directory server operators will
1696 be able to filter out most hostile ORs.
1699 <em>Convince the directories that a malfunctioning OR is
1700 working.</em> In the current Tor implementation, directory servers
1701 assume that an OR is running correctly if they can start a TLS
1702 connection to it. A hostile OR could easily subvert this test by
1703 accepting TLS connections from ORs but ignoring all cells. Directory
1704 servers must actively test ORs by building circuits and streams as
1705 appropriate. The tradeoffs of a similar approach are discussed
1711 try to deny Bob service by flooding his introduction points with
1712 requests. Because the introduction points can block requests that
1713 lack authorization tokens, however, Bob can restrict the volume of
1714 requests he receives, or require a certain amount of computation for
1715 every request he receives.
1719 disrupt a location-hidden service by disabling its introduction
1720 points. But because a service's identity is attached to its public
1721 key, the service can simply re-advertise
1722 itself at a different introduction point. Advertisements can also be
1723 done secretly so that only high-priority clients know the address of
1724 Bob's introduction points or so that different clients know of different
1725 introduction points. This forces the attacker to disable all possible
1726 introduction points.
1730 Bob's introduction point can flood Bob with
1731 introduction requests, or prevent valid introduction requests from
1732 reaching him. Bob can notice a flood, and close the circuit. To notice
1733 blocking of valid requests, however, he should periodically test the
1734 introduction point by sending rendezvous requests and making
1735 sure he receives them.
1739 point is no more sensitive than any other OR on
1740 a circuit, since all data passing through the rendezvous is encrypted
1741 with a session key shared by Alice and Bob.
1747 </a>
1752 matures. (For comparison, the current remailer network
1754 many have 10Mb. The number of users varies (and of course, it's hard to
1755 tell for sure), but we sometimes have several hundred users — administrators at
1756 several companies have begun sending their entire departments' web
1757 traffic through Tor, to block other divisions of
1758 their company from reading their traffic. Tor users have reported using
1759 the network for web browsing, FTP, IRC, AIM, Kazaa, SSH, and
1760 recipient-anonymous email via rendezvous points. One user has anonymously
1761 set up a Wiki as a hidden service, where other users anonymously publish
1762 the addresses of their hidden services.
1766 cells (a bit under half a gigabyte) per week. On average, about 80%
1767 of each 498-byte payload is full for cells going back to the client,
1768 whereas about 40% is full for cells coming from the client. (The difference
1769 arises because most of the network's traffic is web browsing.) Interactive
1770 traffic like SSH brings down the average a lot — once we have more
1771 experience, and assuming we can resolve the anonymity issues, we may
1772 partition traffic into two relay cell sizes: one to handle
1773 bulk traffic and one for interactive traffic.
1776 Based in part on our restrictive default exit policy (we
1777 reject SMTP requests) and our low profile, we have had no abuse
1778 issues since the network was deployed in October
1779 2003. Our slow growth rate gives us time to add features,
1780 resolve bugs, and get a feel for what users actually want from an
1781 anonymity system. Even though having more users would bolster our
1782 anonymity sets, we are not eager to attract the Kazaa or warez
1783 communities — we feel that we must build a reputation for privacy, human
1784 rights, research, and other socially laudable activities.
1787 As for performance, profiling shows that Tor spends almost
1788 all its CPU time in AES, which is fast. Current latency is attributable
1789 to two factors. First, network latency is critical: we are
1790 intentionally bouncing traffic around the world several times. Second,
1791 our end-to-end congestion control algorithm focuses on protecting
1792 volunteer servers from accidental DoS rather than on optimizing
1793 performance. To quantify these effects, we did some informal tests using a network of 4
1797 direct download. We ran a similar test on the production Tor network,
1799 while a direct
1800 download consistently took about 0.3s, the performance through Tor varied.
1803 of building a slow circuit (one that includes a slow or heavily loaded node
1804 or link) is increasing. On the other hand, as our users remain satisfied
1805 with this increased latency, we can address our performance incrementally as we
1806 proceed with development.
1812 Although Tor's clique topology and full-visibility directories present
1813 scaling problems, we still expect the network to support a few hundred
1815 more distributed. With luck, the experience we gain running the current
1816 topology will help us choose among alternatives when the time comes.
1822 </a>
1825 In addition to the non-goals in
1827 before we can be confident of Tor's security.
1830 Many of these open issues are questions of balance. For example,
1831 how often should users rotate to fresh circuits? Frequent rotation
1832 is inefficient, expensive, and may lead to intersection attacks and
1833 predecessor attacks [<a href="#wright03" name="CITEwright03">54</a>], but infrequent rotation makes the
1834 user's traffic linkable. Besides opening fresh circuits, clients can
1835 also exit from the middle of the circuit,
1836 or truncate and re-extend the circuit. More analysis is
1837 needed to determine the proper tradeoff.
1842 How should we choose path lengths? If Alice always uses two hops,
1843 then both ORs can be certain that by colluding they will learn about
1844 Alice and Bob. In our current approach, Alice always chooses at least
1845 three nodes unrelated to herself and her destination.
1846 Should Alice choose a random path length (e.g. from a geometric
1847 distribution) to foil an attacker who
1848 uses timing to learn that he is the fifth hop and thus concludes that
1849 both Alice and the responder are running ORs?
1852 Throughout this paper, we have assumed that end-to-end traffic
1853 confirmation will immediately and automatically defeat a low-latency
1854 anonymity system. Even high-latency anonymity systems can be
1855 vulnerable to end-to-end traffic confirmation, if the traffic volumes
1856 are high enough, and if users' habits are sufficiently
1857 distinct [<a href="#statistical-disclosure" name="CITEstatistical-disclosure">14</a>,<a href="#limits-open" name="CITElimits-open">31</a>]. Can anything be
1858 done to
1859 make low-latency systems resist these attacks as well as high-latency
1860 systems? Tor already makes some effort to conceal the starts and ends of
1861 streams by wrapping long-range control commands in identical-looking
1862 relay cells. Link padding could frustrate passive observers who count
1863 packets; long-range padding could work against observers who own the
1864 first hop in a circuit. But more research remains to find an efficient
1865 and practical approach. Volunteers prefer not to run constant-bandwidth
1866 padding; but no convincing traffic shaping approach has been
1867 specified. Recent work on long-range padding [<a href="#defensive-dropping" name="CITEdefensive-dropping">33</a>]
1868 shows promise. One could also try to reduce correlation in packet timing
1869 by batching and re-ordering packets, but it is unclear whether this could
1870 improve anonymity without introducing so much latency as to render the
1871 network unusable.
1874 A cascade topology may better defend against traffic confirmation by
1875 aggregating users, and making padding and
1876 mixing more affordable. Does the hydra topology (many input nodes,
1877 few output nodes) work better against some adversaries? Are we going
1878 to get a hydra anyway because most nodes will be middleman nodes?
1881 Common wisdom suggests that Alice should run her own OR for best
1882 anonymity, because traffic coming from her node could plausibly have
1883 come from elsewhere. How much mixing does this approach need? Is it
1884 immediately beneficial because of real-world adversaries that can't
1885 observe Alice's router, but can run routers of their own?
1888 To scale to many users, and to prevent an attacker from observing the
1889 whole network, it may be necessary
1890 to support far more servers than Tor currently anticipates.
1891 This introduces several issues. First, if approval by a central set
1892 of directory servers is no longer feasible, what mechanism should be used
1893 to prevent adversaries from signing up many colluding servers? Second,
1894 if clients can no longer have a complete picture of the network,
1895 how can they perform discovery while preventing attackers from
1896 manipulating or exploiting gaps in their knowledge? Third, if there
1897 are too many servers for every server to constantly communicate with
1898 every other, which non-clique topology should the network use?
1899 (Restricted-route topologies promise comparable anonymity with better
1900 scalability [<a href="#danezis-pets03" name="CITEdanezis-pets03">13</a>], but whatever topology we choose, we
1901 need some way to keep attackers from manipulating their position within
1902 it [<a href="#casc-rep" name="CITEcasc-rep">21</a>].) Fourth, if no central authority is tracking
1903 server reliability, how do we stop unreliable servers from making
1904 the network unusable? Fifth, do clients receive so much anonymity
1905 from running their own ORs that we should expect them all to do
1906 so [<a href="#econymics" name="CITEeconymics">1</a>], or do we need another incentive structure to
1907 motivate them? Tarzan and MorphMix present possible solutions.
1912 When a Tor node goes down, all its circuits (and thus streams) must break.
1913 Will users abandon the system because of this brittleness? How well
1915 node failure? If affected users rebuild circuits immediately, how much
1916 anonymity is lost? It seems the problem is even worse in a peer-to-peer
1917 environment — such systems don't yet provide an incentive for peers to
1918 stay connected when they're done retrieving content, so we would expect
1919 a higher churn rate.
1927 </a>
1930 Tor brings together many innovations into a unified deployable system. The
1931 next immediate steps include:
1935 has led us to adopt a clique topology, semi-centralized
1936 directories, and a full-network-visibility model for client
1937 knowledge. These properties will not scale past a few hundred servers.
1939 approaches, but more deployment experience will be helpful in learning
1940 the relative importance of these bottlenecks.
1944 good bandwidth and latency. We should instead adopt the MorphMix model,
1945 where nodes advertise their bandwidth level (DSL, T1, T3), and
1946 Alice avoids bottlenecks by choosing nodes that match or
1947 exceed her bandwidth. In this way DSL users can usefully join the Tor
1948 network.
1952 and possibly better anonymity [<a href="#econymics" name="CITEeconymics">1</a>]. More nodes means increased
1953 scalability, and more users can mean more anonymity. We need to continue
1954 examining the incentive structures for participating in Tor. Further,
1955 we need to explore more approaches to limiting abuse, and understand
1956 why most people don't bother using privacy systems.
1960 in performance and bandwidth are clear but its security benefits are
1961 not well understood. We must pursue more research on link-level cover
1962 traffic and long-range cover traffic to determine whether some simple padding
1963 method offers provable protection against our chosen adversary.
1969 caching web proxy [<a href="#shsm03" name="CITEshsm03">47</a>], to improve anonymity for cached pages
1970 (Alice's request never
1971 leaves the Tor network), to improve speed, and to reduce bandwidth cost.
1972 On the other hand, forward security is weakened because caches
1973 constitute a record of retrieved files. We must find the right
1974 balance between usability and security.
1978 Clients currently download a description of
1979 the entire network every 15 minutes. As the state grows larger
1980 and clients more numerous, we may need a solution in which
1981 clients receive incremental updates to directory state.
1982 More generally, we must find more
1983 scalable yet practical ways to distribute up-to-date snapshots of
1984 network status without introducing new attacks.
1989 external review. We hope that as Tor
1990 is deployed, more people will examine its
1991 specification.
1995 designer of MorphMix to unify the specification and implementation of
1996 the common elements of our two systems. So far, this seems
1997 to be relatively straightforward. Interoperability will allow testing
1998 and direct comparison of the two designs for trust and scalability.
2002 gain experience in deploying an anonymizing overlay network, and
2003 learn from having actual users. We are now at a point in design
2004 and development where we can start deploying a wider network. Once
2005 we have many actual users, we will doubtlessly be better
2006 able to evaluate some of our design decisions, including our
2007 robustness/latency tradeoffs, our performance tradeoffs (including
2008 cell size), our abuse-prevention mechanisms, and
2009 our overall usability.
2016 We thank Peter Palfrader, Geoff Goodell, Adam Shostack, Joseph Sokol-Margolis,
2017 John Bashinski, and Zack Brown
2018 for editing and comments;
2019 Matej Pfajfar, Andrei Serjantov, Marc Rennhard for design discussions;
2020 Bram Cohen for congestion control discussions;
2021 Adam Back for suggesting telescoping circuits; and
2023 This work has been supported by ONR and DARPA.
2035 On the economics of anonymity.
2040 </dd>
2042 R. Anderson.
2043 The eternity service.
2047 </dd>
2049 The Anonymizer.
2053 </tt></dd>
2056 Freedom systems 2.1 security issues and analysis.
2057 White paper, Zero Knowledge Systems, Inc., May 2001.
2060 </dd>
2063 Traffic analysis attacks and trade-offs in anonymity providing
2064 systems.
2069 </dd>
2072 The EAX mode of operation: A two-pass authenticated-encryption
2073 scheme optimized for simplicity and efficiency.
2077 </dd>
2080 Web MIXes: A system for anonymous and unobservable Internet
2081 access.
2083 Technologies: Workshop on Design Issue in Anonymity and Unobservability</em>.
2087 </dd>
2090 Freedom systems 2.0 architecture.
2091 White paper, Zero Knowledge Systems, Inc., December 2000.
2094 </dd>
2096 Z. Brown.
2097 Cebolla: Pragmatic IP Anonymity.
2101 </dd>
2103 D. Chaum.
2104 Untraceable electronic mail, return addresses, and digital
2105 pseudo-nyms.
2109 </dd>
2112 Wide-area cooperative storage with CFS.
2117 </dd>
2119 W. Dai.
2120 Pipenet 1.1.
2121 Usenet post, August 1996.
2126 </tt></dd>
2128 G. Danezis.
2129 Mix-networks with restricted routes.
2134 </dd>
2136 G. Danezis.
2137 Statistical disclosure attacks.
2142 </dd>
2145 Mixminion: Design of a type III anonymous remailer protocol.
2147 IEEE CS, May 2003.
2150 </dd>
2153 Using Client Puzzles to Protect TLS.
2155 Aug. 2001.
2158 </dd>
2161 The TLS Protocol - Version 1.0.
2165 </dd>
2168 A Reputation System to Increase MIX-net Reliability.
2173 </dd>
2176 The free haven project: Distributed anonymous storage service.
2178 Technologies: Workshop on Design Issue in Anonymity and Unobservability</em>.
2182 </dd>
2185 Tor protocol specifications.
2189 </tt></dd>
2192 Reliable MIX Cascade Networks through Reputation.
2197 </dd>
2199 J. Douceur.
2200 The Sybil Attack.
2205 </dd>
2208 MIXes in mobile communication systems: Location management with
2209 privacy.
2214 </dd>
2217 Tarzan: A peer-to-peer anonymizing network layer.
2222 </dd>
2225 Herbivore: A scalable and efficient protocol for anonymous
2226 communication.
2227 Technical Report TR2003-1890, Cornell University Computing and
2228 Information Science, February 2003.
2231 </dd>
2233 I. Goldberg.
2235 PhD thesis, UC Berkeley, Dec 2000.
2238 </dd>
2241 Hiding routing information.
2246 </dd>
2249 Mixing E-mail with Babel.
2254 </dd>
2256 A. Hintz.
2257 Fingerprinting websites using traffic analysis.
2262 </dd>
2264 A. Jerichow, J. Müller, A. Pfitzmann, B. Pfitzmann, and M. Waidner.
2265 Real-time mixes: A bandwidth-efficient anonymity protocol.
2270 </dd>
2273 Limits of anonymity in open environments.
2278 </dd>
2281 SOCKS.
2286 </dd>
2289 Timing analysis in low-latency mix-based systems.
2291 LNCS (forthcoming), 2004.
2294 </dd>
2297 Hordes: A multicast-based protocol for anonymity.
2301 </dd>
2303 C. Meadows.
2304 The NRL protocol analyzer: An overview.
2308 </dd>
2311 Mixmaster Protocol - Version 2.
2312 Draft, July 2003.
2316 </tt></dd>
2319 The Dark Side of the Web: An Open Proxy's View.
2323 </tt></dd>
2326 ISDN-mixes: Untraceable communication with very small bandwidth
2327 overhead.
2332 </dd>
2334 Privoxy.
2338 </tt></dd>
2341 Protocols using anonymous connections: Mobile applications.
2347 </dd>
2350 Anonymous connections and onion routing.
2355 </dd>
2358 Crowds: Anonymity for web transactions.
2362 </dd>
2365 Practical anonymity for the masses with morphmix.
2367 LNCS (forthcoming), 2004.
2370 </dd>
2373 Analysis of an Anonymity Network for Web Browsing.
2378 </dd>
2381 Passive attack analysis for connection-based anonymity systems.
2386 </dd>
2391 CS, 2002.
2394 </dd>
2397 Using caching for browsing anonymity.
2401 </dd>
2404 Onion Routing access configurations.
2405 In <em>DARPA Information Survivability Conference and Exposition
2409 </dd>
2412 Towards an Analysis of Onion Routing Security.
2414 Technologies: Workshop on Design Issue in Anonymity and Unobservability</em>,
2418 </dd>
2420 A. Tannenbaum.
2421 Computer networks, 1996.
2424 </dd>
2426 The AN.ON Project.
2427 German police proceeds against anonymity service.
2428 Press release, September 2003.
2433 </tt></dd>
2436 Tangler: A censorship-resistant publishing system based on document
2437 entanglements.
2442 </dd>
2445 Publius: A robust, tamper-evident, censorship-resistant and
2446 source-anonymous web publishing system.
2448 2000.
2451 </dd>
2454 Defending anonymous communication against passive logging attacks.
2457 </dl>
2464 <a name="tthFtNtAAB"></a><a href="#tthFrefAAB"><sup>1</sup></a>Actually, the negotiated key is used to derive two
2465 symmetric keys: one for each direction.
2468 With 48 bits of digest per cell, the probability of an accidental
2469 collision is far lower than the chance of hardware failure.
2472 Rather than rely on an external infrastructure, the Onion Routing network
2473 can run the lookup service itself. Our current implementation provides a
2474 simple lookup system on the
2475 directory servers.
2478 attack should not be confused with the much more complicated latency
2479 attacks of [<a href="#back01" name="CITEback01">5</a>], which require a fingerprint of the latencies
2480 of all circuits through the network, combined with those from the
2481 network edges to the target user and the responder website.
2482 <br /><br /><hr /><small>File translated from
2487 </body></html>