Predicting Tor Path Compromise by Exit Port Tor Network Client Entry Guard Exit Router Destination Host Middle Router Circuit Directory Server Kevin Bauer, Dirk Grunwald, and Douglas Sicker University of Colorado IEEE WIDA 2009 December 16, 2009 Tor: Anonymity for TCP Applications Last hop knows the destination Colluding entry and exit routers can use simple timing analysis to de-anonymize the client and destination [Serjantov et al., 2003; Levine et al., 2004] First hop knows the client Client Tor Network Entry Guard Exit Router Destination Host Middle Router Circuit Directory Server Router List Tor provides anonymity for TCP by tunneling traffic through a virtual circuit of three Tor routers using layered encryption 1 Prior Attacks Against Tor Last hop knows the destination Prior work showed that the likelihood of circuit compromise in Tor is relatively high [Bauer et al., 2007] First hop knows the client Client Exit Router Tor Network Entry Guard Destination Host Middle Router Circuit Directory Server High BW routers chosen most often Router List 1. Clients choose Tor routers in proportion to their bandwidths 2. Tor routers self-advertise their bandwidth capacities Routers can lie! 2 Our Contribution We observe that the bandwidth available for different applications is not uniformly distributed among exit Tor routers We extend prior work by investigating whether certain applications are more vulnerable to attack than others We hypothesize that traffic destined for ports with little bandwidth is 3 more vulnerable to circuit compromise Talk Outline • Background on path selection in Tor • Experimental setup • Experimental results – Exit bandwidth is not uniformly distributed – Long-lived traffic requires “stable” routers • Toward solutions • Future work • Summary and conclusions 4 Path Selection in Tor • Clients choose Tor routers in proportion to their bandwidth capacities • To reduce the risk of path compromise, Tor clients choose their circuits very carefully • Circuit construction rules Mitigates risk of choosing adversary controlled routers • A router may only be used once per circuit • Only one router per /16 network and two routers per IP address Mitigates the • First router must be an entry guard “predecessor attack” Ensures traffic • The exit router must allow connections to the traffic’s can be delivered destination host and port 5 Path Selection: Exit Policies • Tor allows exit routers to specify their own exit policies • Can be used to help router operators manage risk of abuse [Bauer et al., 2008] • Possible Tor router configurations – Non-exit: Router is not allowed to connect to any (non-Tor) Internet host – Exit: May connect to designated port numbers (and hosts) on the Internet Middle Router Client Exit Router Destination Host Entry Guard 6 Path Selection: Stable Paths • Applications with persistent sessions (SSH, FTP) require special routers that have been alive for a long time • Marked as Stable by the directory servers – Stable router is in the top half of all routers in terms of mean time between failures – Or alive for at least 30 days 7 Experimental Evaluation: Setup • We simulate Tor’s router selection algorithm to study how certain applications may be more vulnerable to circuit compromise • Fuel simulations with real Tor router data from the directory servers (May 31, 2009 snapshot) – 1,444 total routers with 403.3 MB total bandwidth – 770 “stable” routers with 326.9 MB total bandwidth • Simulation details – Generate 10,000 circuits for applications (default port): • FTP (21), SSH (22), Telnet (23), SMTP (25), HTTP (80), POP3 (110), HTTPS (443), Kazaa P2P (1214), BitTorrent tracker (6969), Gnutella P2P (6346), and eDonkey P2P (4661) – Add 6 - 106 malicious routers (10 MB/s BW) and count compromised circuits 8 Experimental Evaluation: Results Fraction of circuits that are compromised for each application’s default exit port 6 routers (with 60 MB) make up 12% of the total bandwidth SMTP (outgoing E-mail) and peer-to-peer file sharing applications are more vulnerable to circuit compromise The number of circuits compromised increases as more malicious routers are injected into the network 9 Exit Bandwidth Distribution is Skewed Distribution of exit bandwidth by default exit port number SMTP and peer-to-peer applications have fewest routers and least amount of exit bandwidth Fraction of circuits that are compromised for each application’s default exit port 10 Long-Lived Traffic Needs “Stable” Routers • Applications with persistent sessions require “stable” routers • Only 770/1,444 routers are Stable Distribution of exit bandwidth by default exit port number Fraction of circuits that are compromised for each application’s default exit port • Slightly higher compromise rate than HTTP/HTTPS/Telnet/POP3 11 Only the Exit Router is Malicious • If only the exit router is malicious, an attacker could still learn significant identifying information – i.e., Login credentials • HTTP – 6 malicious routers: Controls exit router 33.6% of the time – 16 malicious routers: Controls exit router 56.5% of the time • FTP – 6 malicious routers: Controls exit router 46.7% of the time – 16 malicious routers: Controls exit router 70.7% of the time • This is a very real threat, since many popular websites still do not provide TLS-protected logins 12 Toward Solutions • One solution is to give users the ability to manage their risk of attack • Prior work proposed that users tune the router selection between bandwidth-weighted and uniform router selection [Snader and Borisov, 2008] – Allows users to trade-off between strong anonymity and strong performance Only 0.09% of BitTorrent tracker Uniform router selection: circuits compromised c > 1 malicious routers E > 0 is number exit routers N > 1 number total routers Compare to 18.5% • However, it remains necessary to balance the traffic load over the available bandwidth • General solutions to this attack is an open problem 13 Future Work: Selective DoS Attacks • Extend this work to consider selective denial-of-service attacks – Attack strategy: If an adversary does not control the endpoints of a given circuit, they disrupt the circuit, causing it to be rebuilt Fraction of circuits that are compromised for each application’s default exit port Initial results with selective denial-of-service Effects of bandwidth disparities are magnified SMTP and peer-to-peer applications show extremely high compromise rate (68-93%) with only 6 malicious routers 14 Summary and Conclusions Tor Network Client Entry Guard Exit Router Destination Host Middle Router Circuit Directory Server • We demonstrated our hypothesis that certain applications are more vulnerable than others to circuit compromise in Tor • Through a simulation study driven by data obtained from the real Tor network, we found that SMTP and peer-to-peer file sharing applications are most vulnerable • We suggest that concerned users tune the router selection bias to control the risk of path compromise 15