- tor-dev - lists.torproject.org

Blocking-Test: GSoC Penultimate Status Report
by Brandon Wiley 12 Aug '11

12 Aug '11

Scoring and reporting for detectors and encoders is pretty much done. The graphs aren't quite ready yet, for aesthetic reasons, but I have some nice tables to report: Length detector: 16% accuracy Timing detector: 89% accuracy Entropy detector: 94% accuracy SSL was correctly identified 25% of the time, 70% of the time other protocols were misidentified as being SSL, and 5% of the time SSL was misidentified as another protocol. obfsproxy was correctly identified 55% of the time, 10% of the time other protocols were misidentified as being obfsproxy, and 35% of the time obfsproxy was misidentified as another protocol. Dust was correctly identified 48% of the time, 4% of the time other protocols were misidentified as being Dust, and 48% of the time Dust was misidentified as another protocol. obfsproxy is distinguishable from SSL 96% of the time. obfsproxy is distinguishable from Dust 98% of the time. Dust is distinguishable from SSL 56% of the time. I'm sure these numbers are confusing as several different sorts of things are being measured, all with similar numbers. It is probably quite unclear how all of these percentages add up. I think this is where the graphs will help. In particular I think it's hard to visualize accuracy when there are both false positives and false negatives to take into account. I think the most interesting result here is the entropy detector. It's the simplest and also the most effective. Not only that, but the entropy detector only looks at the entropy of the first packet, so it's inexpensive. The original reason that I implemented this detector was that there was an argument against Dust that it could be trivially filtered by setting an entropy threshold above which all traffic is filtered. Surprisingly, Dust has lower entropy than SSL, so this attack will not work to specifically target Dust. This is unexpected because Dust uses totally random bytes, whereas the SSL handshake does not. This is a result of a general issue with how Shannon entropy is defined. Shannon entropy is normally defined across a large statistical sampling, in which case all three protocols tested will converge to maximum since they are all encrypted and therefore apparently random. However, the attacker does not in this case get a large statistical sample. Since they don't know a priori which connections are Dust connections, they only get however many packets are in each trace to use as samples. The current version of Dust that I'm testing has most of the advanced obfuscation stripped out as I haven't had time to add it back in after completely reimplementing the protocol to use TCP instead of UDP and all of the modified assumptions that come with that change. So this version of Dust is basically just the ntor protocol with no packet length or timing hiding in use. The first packet is therefore 32 random bytes consisting of the curve25519 ephemeral ECDH key, as compared to the 1k first packet of SSL. This is the reason for the low entropy calculation, because given only 32 bytes to sample almost all possible bytes are going to have a probability of 0 or 1. There is an upper bound on entropy when sample sizes are small. Achieving maximum first-order entropy here would require at least 512 bytes so that each value 0 to 255 could occur exactly twice. There are of course many ways to measure entropy and you could say that I'm just doing it wrong. Some alternative measurements have already been suggested that I might implement as well. I am not suggesting that Dust is in any way immune to entropy-based detection. In fact, Dust is currently detected very well by the current entropy detector (it passed 100% of the packet length and timing tests, so all detection of Dust was done by the entropy detector) precisely because it has strangely low entropy because of it's small first packets. I bring this up only as an interesting example of how intuition can fail when trying to use Shannon entropy to measure individual messages instead of channels. As far as the state of the project goes, I think the project can be considered done. Looking at the project documentation<http://stepthreeprivacy.org/documentation>, everything didn't work out exactly as originally planned, I think it all worked out for the best and the project is where I wanted it to be by the completion of Summer of Code. I have generation of HTTP and HTTPS traffic, both over Tor and plain. I have two encoders: obfs2 and Dust. I have three detectors: length, timing, and entropy. The string detector didn't work out, but the entropy detector worked out better than planned, so on balance I think it was a win. I didn't implement all of the detector modes because upon further reflection it doesn't really make sense to just try all of these different modes. You want the mode that works best. So for packet lengths and timing that is full conversations and for entropy that's the first packet (could be first N packets, but I tried N=1 and it worked great). Random sampling of packets just seems like a way to limit your success at this point when non-random sampling of just the first packet seems to be so effective. All of the utilities have been written, and more. I have the specified utilities for separating the pcap files into streams, and extracting the strings, lengths, and timings. Additionally, I have some utilities for extracting the entropy of streams, tagging the streams with what protocol they are using (by port), and extracting dictionary and corpus models for latent semantic analysis. I also have some utilities for graphing distributions of properties of different protocols which I hope to finish up next week during the wrap-up period. Looking to the future, I have changed my goals somewhat from what I had in mind at the beginning of the project. After having attended the TorDev meetup in Waterloo and the PETS conference, I am much more interested in integrating my work with the Tor project instead of just testing the various protocols that have been proposed in order to see how my own protocol stacks up against the competition. As part of this transition, I have rewritten Dust from scratch, both in terms of implementation and in terms of the protocol, in order to be better suited for use as a pluggable transport for Tor. My goal for Dust is to have it shipped with both Tor clients and bridges so that it can be used in the field to contact bridges despite DPI filtering of Tor connections. My goal now for the blocking-test project is to focus on blocking resistance for Tor. So instead of adding all of the protocols under the sun, I'm going to concentrate on protocols that have been implemented or are under consideration to become pluggable transports. Additionally, the next protocol I'm going to add after GSoC is over is just plain Tor. There is at least one bug filed about a suspicion that Tor can currently be fingerprinted based on some packet characteristics and I'd like to be able to close that bug because we now have the capability to easily generate the necessary information. Finally, I would like to offer some insight into what I think is the future of blocking-resistant transports now that I have done some work trying to break them. Fundamentally, the only thing an attacker can do is limit the bitrate of the connection. Given any sort of channel, we can encode arbitrary information over that channel. However, the smaller the channel the slower the bitrate. The goal of encodings should be to maximize the bitrate given the constraints of the censor. It's easy to develop very slow encodings such as natural language encodings over HTTP. However, in order to maximize the throughput you need to have a good definition of the attack model. All encoding for blocking-resistant purposes lowers the efficiency above just transmitting the normal content (assumed it has already been compressed). Therefore all encoded conversations are going to be longer than unencoded conversations. I think this is the future of writing detectors because it's a fundamental constraint on encoding. I think this should even be able to defeat something like Telex (although let's not get into the details of attacks on Telex specifically, I'm talking about a general idea here). Given a (static, for the sake of argument simplicity) website, I can download every page on that website. Then I can compare the length of intercepted downloads and detect when they don't match up. For dynamic websites you can do the same thing with a statistical model. I think, therefore, that the future of blocking-resisting protocols is to encode single logical connections over an ensemble of multiple actual connections. Ideally these connections would mimic normal traffic in terms of number of connections, duration, directional flow rates, etc.. It's something to think about anyway. It could be an interesting problem.

1 0

Instructions for building the Vidalia-bundle?
by Steve Snyder 10 Aug '11

10 Aug '11

Is there documentation anywhere on how to build the Vidalia-bundle for Windows? If so, where can I find it? Thanks.

3 3

GSoC Update
by Brandon Wiley 06 Aug '11

06 Aug '11

With the TorDev meetup and everything it's been a while since I sent out an update on the state of the project, although I think we more or less caught up on the status at the meetup. The bulk of the implementation work had already been completed, so what I've been working on since then is just cleaning up the code and rewriting some parts. In particular this week I rewrote the detectors to be mostly python. The core Bayesian model is still in BUGS/JAGS and there is still R code to interface with that. However, the output of the models after training is now converted to a CSV file and the actually detection code is a python program which reads the CSV file with the trained model's prediction data and the compares it to the data of a stream to be classified. This means that you only need R for training the models. If we can put together a canonical model for each protocol then you can use it to classify new traffic just using python. I think this is a win because R was the bulkiest dependency. It also segments the process more nicely in that there can be some people gathering traces, other people building models, and then other people running detectors against new traces. In terms of code cleanup, I've been adding comments and refactoring large files (in particular pavement.py) to be more readable. I also have the scoring tasks implemented so that, given the output of the detectors, the various detectors and encoders are scored based on their number of correct guesses, false positives, and false negatives. The only functionality left to implement is some sort of reporting, probably in the form of graphs of the scores so that detectors and encoders can be compared. The rest of the work is just going to be refining everything and getting both the workflow and the code structure to be smooth. Also, I've stopped working on the string detector because I realized that it's irrelevant for the two encoders I actually have support for right now: obfs2 and Dust. They both already use totally randomized packet contents and so will not be vulnerable to this kind of detector. Other encodings will be vulnerable, but the string detector can wait until we support these encodings. So instead I've implemented an entropy detector, which should work very well against both obfs2 and the current implementation of Dust. The entropy detector is actually the easiest model so far as the entropy of a trace is just a single number, so that's convenient.

1 0

Requesting feedback on TorDNSd v1.1
by LETO 02 Aug '11

02 Aug '11

Could some of you be so kind to try out my latest version of TorDNSd? It works a lot like ttdnsd with some additional features: - Filters to tell TorDNSd what requests to forward directly, forward through the SOCKS proxy or reject. - Remaps to define your own query replies (Currently only supports A, MX and NS replies) - Interactive shell (tordnsd-shell.exe): config-add settings and see them in action right away, will add extra features in the future. - Accepts both UDP and TPC DNS requests, forwards both through TCP requests - Basic caching of replies (optional) The rejects are handy to for example prevent leaking .onion requests (or any request if you want) You can also find some additional information about this release on my blog ( http://leto-r.blogspot.com/2011/07/tordnsd-v11-is-out.html ) It is written in C# and has been tested by myself under mono 2.6.7 / ubuntu natty. Check the default configuration @ https://raw.github.com/LETO-R/TorDNSd/b7aa04e980ad62308d4f2fa3143700c85b537… the full list of available settings. You can either download the mono compatible bins ( https://github.com/downloads/LETO-R/TorDNSd/tordnsd-v1.1-all-bin.zip ) or the windows installer ( https://github.com/downloads/LETO-R/TorDNSd/tordnsd-v1.1-win-setup.exe ) You can find the source code of these bins @ https://github.com/LETO-R/TorDNSd/tree/b7aa04e980ad62308d4f2fa3143700c85b53… sure to get the submodules too) The sourcecode should be compilable using either MonoDevelop 2.4.2 (using the mono runtime) or Visual Studio 2010. If you do not run it as root on a non-windows system, a fatal error is printed since TorDNSd will most likely not be able to bind on port 53. Specify --no-root to skip this check. Kind of feedback I'd like of you: - What OS did you try it on? When ran using mono, what version? - Did it work? Did it require the elevated rights (root / administrator)? Did the root-checking work on non-windows systems? - Did the shell work for you? If not, what issues did you notice? Any improvements that you suggest? - Are the settings explained well enough (check tordnsd.conf)? - Is the supplied default configuration sufficient? Any filter / remaps I forgot? - Any other feedback you can think of, don't hold back! This is my first attempt at a cross-platform .NET project that contains some non-windows specific code / workarounds (mostly for the shell) so I really could use the feedback. If you do not wish to run it as your main dns resolver 'just yet', you can always test it by using dig (examples: 'dig @127.0.0.1 bla.onion' to test rejection, or 'dig @127.0.0.1 vescum.tor' to test the remap feature) All I ask is to stay constructive. If you wonder, I constantly have TorDNSd (shell) running now and have configured it as my main dns resolver as I consider it secure. Hope you find this to be true as well. - LETO

3 5

[Patch] or/networkstatus.c
by Gisle Vanem 01 Aug '11

01 Aug '11

Hi list. In or/networkstatus.c there is a "#if 0" block inside the macro "SMARTLIST_FOREACH_JOIN". Not all compilers handle such contructs. In the prosess of making solution/projects file for "MS Visual C++ 2010 Express", I hit this problem (cl Version 16.00.30319.01). Can you please accept the following patch: diff --git a/src/or/networkstatus.c b/src/or/networkstatus.c index 2586ce6..d8d6680 100644 --- a/src/or/networkstatus.c +++ b/src/or/networkstatus.c @@ -2001,19 +2001,25 @@ routers_update_status_from_consensus_networkstatus(smartlist_t *routers, routers_sort_by_identity(routers); + /* Since not all compilers handles '#' inside macros, we use a helper-macro. */ +#if 0 + #define CHECK_ROUTER_PURPOSE() do { \ + /* We have no routerstatus for this router. Clear flags and skip it. */ \ + if (!authdir) { \ + if (router->purpose == ROUTER_PURPOSE_GENERAL) \ + router_clear_status_flags(router); \ + } \ + } while (0) +#else + #define CHECK_ROUTER_PURPOSE() +#endif + SMARTLIST_FOREACH_JOIN(ns->routerstatus_list, routerstatus_t *, rs, routers, routerinfo_t *, router, tor_memcmp(rs->identity_digest, router->cache_info.identity_digest, DIGEST_LEN), + { CHECK_ROUTER_PURPOSE(); }) { -#if 0 - /* We have no routerstatus for this router. Clear flags and skip it. */ - if (!authdir) { - if (router->purpose == ROUTER_PURPOSE_GENERAL) - router_clear_status_flags(router); - } -#endif - }) { /* We have a routerstatus for this router. */ const char *digest = router->cache_info.identity_digest; ------------------------ --gv

2 2

Unused -v and -F options in torperf/trivsocks-client?
by David Fifield 30 Jul '11

30 Jul '11

Running "make" in torperf, I get errors about unused variables: $ make gcc -Wall -Werror -ggdb -c trivsocks-client.c trivsocks-client.c: In function ‘main’: trivsocks-client.c:452:51: error: variable ‘force’ set but not used [-Werror=unused-but-set-variable] trivsocks-client.c:452:21: error: variable ‘isVerbose’ set but not used [-Werror=unused-but-set-variable] cc1: all warnings being treated as errors make: *** [trivsocks-client.o] Error 1 The variables have to do with the -v and -F options, which appear not to do anything. This patch removes the options, but is it intentional? David Fifield

3 4

Improving Private Browsing Mode/Tor Browser
by Georg Koppen 29 Jul '11

29 Jul '11

After reading Mike's blog post and the material contained in it (via links) I thought it would be helpful to start a discussion about it. First of all thanks for explaining the idea of improving the private browsing mode. That aim seems worthwile but I want to focus more on the needs for high anonymity systems like Tor (I am one of the JonDos people to set the records straight). Thus, first I am not sure about the relationship of the improved private browsing and the anon mode. It seems like the former is kind of precondition of the latter and the latter adds some special anon features (or just layout stuff??): "We would love to be able to ship a vastly simplified browser extension that contains only a compiled Tor binary and some minimal addon code that simply "upgrades" the user's private browsing mode into a fully functional anonymous mode." What shall the upgrade do exactly? And why having again add-ons that can probably be toggled on/off and are thus more error-prone than just having an, say, Tor anon mode? Or is this already included in the Tor anon mode but only separated in the blog post for explanatory purposes? Sticking to the blog post (one of) its central idea seems to be to isolate the identifiers and state to the top-level domain in the URL bar as "activity in Tor Browser on one site should not trivially de-anonymize their activity [i.e. the activity of Tor users, G.K.] on another site to ad networks and exits". I am wondering whether this idea really helps here at least regarding exit mixes. If one user requests google.com, mail.google.com and other Google services within the 10 minutes interval (I am simplifying here a bit) without deploying TLS the exit is still able to connect the whole activity and "sees" which services that particular user is requesting/using. Even worse, if the browser session is quite long there is a chance of recognizing that user again if she happens to have the same exit mix more than once. Thus, I do not see how that helps avoiding linkability for users that need/want strong anonymity while surfing the web. Would be good to get that explained in some detail. Or maybe I am missing a point here. Now something to the proposed behavior of the referer and window.name. It is said that they should adhere to the "same-origin policy where sites from different origins get either no referer, or a referer that is truncated to the top-level domain". Assuming I understood TorButton's Smart-Spoofing option properly: Why is it not applied to the referer/window.name anymore? In other words: Why is the referer (and window.name) not kept if the user surfs within one domain (let's say from example.com to foo.example.com and then to foo.bar.example.com)? Before we implemented almost the same algorithm than Torbutton's smart-spoof algo in our own extension a while ago we had some compatibility issues (I remember yahoo.com that needed to have the referer untouched while surfing within the domain) that got fixed by it and never popped up again. Why stepping back? The idea "sites should also be allowed to request an exemption to this rule on a per-site basis using an html attribute, which could trigger a chrome permissions request, or simply be granted automatically (on the assumption that they could just URL smuggle the data)" seems rather awkward and not a good solution to a problem that is not really one. One final point: Leaving my previous section aside: Why is the referer and window.name not treated in the same way as cookies and others in the proposal. Why having two different policies for identifiers? I am not sure whether there could emerge some attacks out of that distinction but my feeling tells me that there should ideally be just one policy that governs all those identifiers. At least it would probably be easier to implement and to audit them. Georg

3 20

Report from the "Tor ecosystem" BoF at DebConf 11
by Jérémy Bobbio 29 Jul '11

29 Jul '11

Hi! Here is a report from DebConf 11, the annual conference of the Debian project. We had the opportunity to hold a "Tor ecosystem" BoF. It was short (we had only 45 minutes), but productive from my perspective. Here's an (obviously partial) report. A lot more people showed up (± 25) than we expected for a pure work meeting, so there is definitely an interest in having Tor and its ecosystem working well in Debian. Among the attendees we had: * Peter Palfrader, maintainer of the 'tor' package, * Ulises Vitulli, (co-)maintainer of 'vidalia', 'tor-arm', 'python-torctl', 'torchat', * Jérémy Bobbio, maintainer of 'torbutton', * intrigeri and bertagaz, developers of Tails (which is a live system based on Debian). A summary follows from some notes and memory. Feel free to correct me if you feel that things happened differently. New upstream packaging policy ============================= We took some time to discuss the changes to the packaging policy that were proposed by Andrew Lewman last June. See <https://lists.torproject.org/pipermail/tor-dev/2011-June/002780.html> Quoting the relevant part: GNU/Linux (deb and rpm) 1. tor-client. The same TBB we ship today. 2. tor-relay. Create tor-relay.(deb|rpm) that includes geoipdb and a pre-configured torrc that is for a non-exit relay. 3. tor-exit-relay. Create tor-exit-relay.(deb|rpm) that includes geoipdb and a pre-configured torrc that is for an exit relay. 4. tor-bridge. Create tor-bridge.(deb|rpm) that includes geoipdb and a pre-configured torrc that is for a bridge. The 'tor' source package in Debian already builds a binary package called 'tor-geoipdb' which contains GeoIP information. We have not even discussed about changing, as it is how it is supposed to be in Debian (pure data package as separate, architecture-independent packages). The proposed 'tor-client' and TBB in general was discussed later. For the rest, the hard part of the discussion also revolved around defining who are target users of these packages. Pure desktop users? People that can fiddle with a command-line and edit a configuration file? Experienced sysadmins? Having 3 different "configuration" packages for relays, exit-relays and bridges is something that could be done in Debian. But we had a rough consensus that it would actually complicate things for most Debian users. For bridges, pure desktop users would probably benefit from a better integration with Vidalia; so that made us rule out the need for a dedicated 'tor-bridge' package. Which left us with the "relay" and "exit relay" use cases and it looked like those could be better adressed by either a debconf question, or by some examples better integrated. Integration with other software, derivatives and custom distributions ===================================================================== While discussing how we could actually make it easier to setup relays, we got back to the lack of a way to source extra configuration files in `torrc`. We outlined some benefits that would have a `/etc/tor/torrc.d` directory full of configuration snippets: * The 'tor' package will not (or less) need patches to change the default settings. * It would make it much easier to implement example setups (relays, exit relays, bridges) either through a debconf question or through a system script like 'tor-setup-relay'. * Tails developers will only need to drop a very small extra configuration file instead of having to rewrite the whole, which should ease development and reviewing. * Other packages which need a hidden service to work (e.g. 'onioncat') would be able to create it automatically by installing the proper configuration snippet. That would enable them to work right after being installed. Feature request: https://trac.torproject.org/projects/tor/ticket/2863 Control of the system-wide 'tor' daemon ======================================= We finally have (currently in 'experimental') an easy way for applications to control the system-wide 'tor' daemon: the ControlSocket option is on by default. Vidalia, ARM and others now only require users to be in the system 'debian-tor' group to be able to control Tor. We did not have the time to properly discuss if, and how, we offer an easy way to add users to this group. Tor Browser Bundle (e.g tor-client) =================================== There is an incentive to have something as close as the Tor Browser Bundle in Debian. Otherwise, desktop users on Debian get a smaller anonymity set (such strong partitioning already happened in the past, see <http://bugs.debian.org/595375>). The main issue is the Tor fork of Firefox. There are a few possibilities on how it can be handled within Debian. Summary: this is doable, but requires a lot of work, and the cooperation of quite a few different parties. One other thing that TBB is doing is to start its own instance of Tor. We agreed that it was a behaviour that would be quite desirable to have. It will make setup even more easier and it might limit users exposure when they are not using Tor. In order to achieve that, we need either to make the 'tor' package not start Tor after being installed or we need a way have the `tor` binary without the rest of the daemon infrastructure (e.g initscript). The former will probably be quite confusing to sysadmins, so we agreed on splitting out the daemon binary from the current 'tor' package to a new 'tor-bin' package. How TorBrowser could get in Debian? =================================== Having a full separate source package for TorBrowser is not possible due to Debian Policy 4.13. But there are still other options, all having their shortcomings: 1. Make the 'iceweasel' source package (that's how Firefox is called in Debian due to trademark issues) produce an extra 'tor-browser' binary package. Issues: - Mike Hommey (iceweasel maintainer) stated clearly that he did not want to maintain the Tor patches. But we can probably settle on a policy like "if it does not build, just drop it". The removal of a binary package does not trigger NEW queue processing. It is also not triggered if the same binary packages returns before two weeks. - Double (?) the build time of 'iceweasel' which is already huge. Security team will not be happy, and it's going to be harder to actually maintain the TorBrowser side of things. Pros: until the patches do not apply anymore, it'll be less maintainance from our side. 2. Make the 'iceweasel' source package produce an extra binary package (e.g 'iceweasel-src') that will contain the full Iceweasel source. Create a 'tor-browser' package that Build-Depends on the former. Issues: - 'iceweasel' is a pretty fast moving target. This will need us to follow its uploads quite closely to ask for binNMU or to upload updated versions quickly. - The security team will have two packages to track instead of one. They might not like it at all. Pros: much easier for Mike Hommey and we are autonomous to upload the 'tor-browser' package if changes are only needed there. Looks like the latter is probably better. But it definitely needs to be checked with all parties involved first. For browser extensions curretly shipped with TBB, Debian already has 'torbutton' and 'noscript'. But we miss 'httpseverwhere' (#591579) and 'betterprivacy'. There are a few open issues that still need to be figured out: * TorBrowser should create dedicated profiles in the user home directories, * How can default values be set? Do we need 'locked down' preferences? * How should we handle other system-wide extensions? Active like for the main Iceweasel, totally ignored, or disabled by default? * If Debian ships the TorBrowser, should we prevent Torbutton from appearing in the usual Iceweasel? This will require a lot of work for the initial bootstrap and a decent amount of work for the daily maintainance after that. But it looked worthwhile anyhow… Cheers, -- Jérémy Bobbio .''`. lunar(a)debian.org : :Ⓐ : # apt-get install anarchism `. `'` `-

1 0

New mailing list for wiki changes
by Karsten Loesing 25 Jul '11

25 Jul '11

Hi everyone, yesterday we found that a few people run their own rss-to-email service to be notified of wiki changes. We decided that it's easier to set up a new mailing list that has all the wiki changes and that everyone can subscribe to. If you're interested in learning when something in the wiki changed, subscribe to: https://lists.torproject.org/cgi-bin/mailman/listinfo/tor-wiki-changes (And if you have suggestions to make the r2e output more useful, please say so. The current configuration is simply 'r2e add "https://trac.torproject.org/projects/tor/timeline?wiki=on&max=50&daysback=2…" tor-wiki-changes(a)lists.torproject.org'.) Best, Karsten

1 0

New Paper: Cloud-based Onion Routing
by Nick Jones 15 Jul '11

15 Jul '11

Hi All, I'm a graduate student at Princeton, and our research group has recently submitted a paper proposing a design for cloud based onion routing. The goal of our research is to securely perform onion routing on cloud based infrastructure (like Amazon EC2 and Rackspace) while allowing users to retain the same (or almost the same) privacy as when using Tor. We distribute trust across multiple cloud providers, and use Chaum's e-cash for payment and access control. Additionally, we hope that the elasticity of cloud infrastructure will make cloud based OR more censorship resistant than current systems. This project is still in a relatively early stage, and we would love to get feedback from the Tor community. We would welcome any comments/questions/criticisms. Our project's website is available at: http://sns.cs.princeton.edu/projects/cor/ A direct link to our paper is here: http://www.cs.princeton.edu/~najones/publications/cor-foci11.pdf Our abstract: Internet censorship and surveillance have made anonymity tools increasingly critical for free and open Internet access. Tor, and its associated ecosystem of vol- unteer traffic relays, provides one of the most secure and widely-available means for achieving Internet anonymity today. Unfortunately, Tor has limitations, including poor performance, inadequate capacity, and a susceptibility to wholesale blocking. Rather than utilizing a large number of volunteers (as Tor does), we propose mov- ing onion-routing services to the “cloud” to leverage the large capacities, robust connectivity, and economies of scale inherent to commercial datacenters. This paper de- scribes Cloud-based Onion Routing (COR), which builds onion-routed tunnels over multiple anonymity service providers and through multiple cloud hosting providers, dividing trust while forcing censors to incur large collat- eral damage. We discuss the new security policies and mechanisms needed for such a provider-based ecosys- tem, and present some preliminary benchmarks. At to- day’s prices, a user could gain fast, anonymous network access through COR for only pennies per day. Thanks! Nick Jones

3 5