PETS in real time: Anonymous communications
27 July 2011
An Accurate System-Wide Anonymity Metric for Probabilistic Attacks
Rajiv Bagai, Huabo Lu, Rong Li, and Bin Tang (Wichita State University)
Traditional entropy based anonymity metrics look at the security of single messages. But how can you quantify the security provided by a whole system? The first paper in this session looks at a system-wide definition of anonymity by “counting” the possible number of matchings between inputs and outputs of an anonymity system. Furthermore, the metric extends to the probabilities over perfect matchings to express subtleties of modern anonymity systems. The paper first of all provides a thorough critique of the metric by Edman et al. (there was also previous work on this metric by the Leuven crew).
In a nutshell the proposed system-wide metric associates a probability to each possible matching, and computes the entropy over this distribution as a measure of anonymity (normalized). The choice of shanon entropy to summarise quality can be changed to min-entropy or other (which is very cool!) One key issue with system-wide metrics is that how they express the properties that any individual message receives. Paul Syverson points out that these type of metrics express more the anonymity capacity of a system — namely how much anonymity the system could provide as a whole. The question of how this capacity for protection is distributed across users may need an extension to those metrics. For anyone who would like to extend metrics to capture this aspect, this paper is a very solid foundation.
DefenestraTor: Throwing out Windows in Tor
Mashael AlSabah, Kevin Bauer and Ian Goldberg (University of Waterloo), Dirk Grunwald (University of Colorado), and Damon McCoy, Stefan Savage, and Geoffrey Voelker (University of California-San Diego)
This paper looks at performance issues within the Tor network, and in particular the effects of the congestion and flow control protocols. Tor implements simple end-to-end flow control mechanism at the granularity of circuits and streams. It turns out that the implemented window based flow control has detrimental effects on performance: it does not protect intermediate routers (who are likely to be the congested ones) from congestion.
Two approaches were followed to solve this problem. First, a smaller window could be used — but this would not solve the problem; or windows can be computed dynamically. Second, the N23 congestion control protocol (used for ATM) could be used over Tor. N23 is simple and guarantees no packets are dropped, while implementing a steady flow of data. Its a credit based system, where packets are sent when credits are available (and consume them), and credits are sent up the network when bandwidth is available.
The evaluation was done under realistic conditions on ExperimenTor. The improvement over the current Tor strategy is significant when it comes to the time to get the first byte, but the time to complete larger (bulk) downloads do suffer (which is part of the point of the protocol).
I am really happy to see research on the intersection of traditional networking and anonymous communications. I have never heard of N23 before (shame on me!), and it seems that it is a good fit for the problem of congestion in anonymity networks (where reliability is not an issue when TCP is used).
Privacy Implications of Performance-Based Peer Selection by Onion Routers: A Real-World Case Study using I2P
Michael Herrmann and Christian Grothoff (Technische Universität München)
This is an attack paper on the I2P network, and in particular the performance based peer selection. It combines a denial-of-service attack to influence the selection of peers within the network, and force a victim to choose corrupt servers.
This is a cute attack that combines denial-of-service, traffic analysis for confirmation you are on the same circuit, and interactions with an infrastructure to attack. This is a very good reminder that anonymity engineering is not simply systems’ work. Every design choice about performance can affect security in dramatic ways. The evaluation was also very sensitive to protecting users: the researchers tried their attack on the real network, but targeted their own circuits (I still want to see details to make sure no other users were affected).
Tor too implements circuit selection on the basis of performance — I am wondering to what extent similar ideas could be applied there …