Protecting Tor from Sybil attacks

Writing  •  Data  •  Code  •  Contact

This research project attempted to identify and characterise Sybil relays in the Tor network. Counting millions of daily users, the Tor network is the largest deployed anonymity network. While the code is developed by The Tor Project, the network of 7,000 relays is operated by volunteers. Sybils are a set of these relays that are controlled by a single operator. While Sybils are not necessarily malicious, they can assist an attacker in (i) manipulating the network’s distributed hash table, (ii) tamper with the network’s exit traffic, or (iii) end-to-end correlate Tor traffic. Therefore, it is important to identify Sybil groups and remove them from the network, if they turn out to be malicious. In this project, we developed techniques to expose Sybil relays, and study the ones we have already discovered. To this end, we have developed sybilhunter—a command line tool to analyse archived Tor network data for signs of Sybil relays. We are using sybilhunter to analyse large piles of data and generate visualisations, such as the ones shown below.

Writing

Data

Relay uptime images

Each hour, Tor’s directory authorities publish the network consensus—an authoritative list of approximately 7,000 Tor relays that are currently running. By analysing archived network consensuses, we are able to visualise the uptime of every single Tor relay that was ever online since 2007. In particular, we create bitmaps whose columns represent relays and whose rows represent consensuses. Each pixel shows the online status of a particular relay at a particular hour. A black pixel means that the relay was online and a white pixel means that the relay was offline. Red blocks highlight suspiciously similar uptime patterns. The image below shows an example—an uptime image for a subset of all relays in January 2014. Note the red blocks; their highly correlated uptime could mean that they are Sybils, run by the same person. We created and uploaded monthly uptime dating back to October 2007, all of which are available online. You can create your own uptime images by using sybilhunter’s “-uptime” command line option.

January 2014 uptime image
Relay uptime image for a subset of January 2014.

Relay churn

The churn rate of a distributed system captures the rate of joining and leaving network participants—in our case Tor relays. We determine the churn rate between two consensuses, published at time t-1 and t, by determining the number of relays that have joined and left the network, respectively. We normalise this number by the consensus size, which gives us a value in the interval [0, 1]. We determined the churn rate between two subsequent network consensuses for all consensuses published since 2007, which leaves us with a time series of churn values. For each relay flag (e.g., Guard, Exit, and HSDir), we create a separate time series. The diagram below shows an example, several days in February 2013, featuring a curious pattern on Monday and Tuesday. Monthly time series diagrams are available online.

February 2013 churn time series
Churn time series for a subset of February 2013.

We also publish our raw data. You can download the following file containing all churn values since 2007, or create your own dataset using sybilhunter’s “-churn” parameter.

Observed fingerprints

We analysed monthly batches of consensuses to determine how many fingerprints each Tor relay—identified by its IP address—used. Relays that change their fingerprints frequently could be misconfigured, or make an attempt to manipulate Tor’s distributed hash table. Below is an example output of this analysis technique, showing four fingerprints that were used by a relay over January 2016.

94.242.206.35 (4 unique fingerprints)
    0760C0A34CCBCB2614075E27B88151FCC7D047EC (seen 603 times)
    BFD74D651690324DCA9FB744E2D54DDC814DD3FB (seen 603 times)
    667C297D3EC6E1281D68F7F4C8C9BE8324D132A3 (seen 120 times)
    0ED2D734F295427E5A3719FA7B9985C335839123 (seen 120 times)

We publish all monthly fingerprint statistics. You can create your own statistics using sybilhunter’s “-fingerprint” argument.

Sybil groups

We created an archive that contains the raw data for the Sybil groups we discuss in our research paper. You can use this archive to verify our results or extend our work. The archive contains a README.txt file that explains the directory structure.

Code

A lot of engineering effort went into sybilhunter, our command line tool to expose and analyse Sybil relays. The tool is implemented in Go and uses our parsing library zoossh. Sybilhunter takes as input archived network data from CollecTor and produces visualisations and CSV files, as the ones shown above. To get sybilhunter, you can download a tarball, or fetch the most recent version over the command line:

go get github.com/NullHypothesis/sybilhunter

Sybilhunter’s README file explains its usage.

For convenience, we reimplemented nearest neighbour ranking—a subset of sybilhunter’s functionality—in Python. The tool is called trnnr (short for Tor relay nearest neighbour ranking) and is available online as well:

git clone https://github.com/NullHypothesis/trnnr

Running the tool is as simple as passing it a relay fingerprint, as its README shows.

Contact

This research was supported in part by the Center for Information Technology Policy at Princeton University. If you have any questions or feedback, please don’t hesitate to get in touch with us.


Last update: 2016-09-09