Inference and security
(George Danezis)
Over the past year at MSR, George has taken advantage of the local expertise in Machine Learning and Bayesian Inference, and tried to apply them to security applications. If successful, this will have a serious impact on security disciplines like traffic analysis, intrusion detection, side channel analysis and steganalysis.
The first completed project relates the problem of detecting Sybil nodes, controlled by an adversary, within a social graph. We show that given a small set of assumptions on the social structure of the graph, we can label each node with the probability they are part of a Sybil attack or not. We experimented on synthetic networks as well as sampled networks from DBLP and LiveJournal. We employed an intern, Prateek Mittal from UIUC to work on this over the summer, and our joint work has just been accepted at NDSS.
George Danezis and Prateek Mittal. SybilInfer: Detecting Sybil Nodes using Social Networks.To appear at NDSS 2009.
The current project, involving our intern from KU Leuven Carmela Troncoso, casts classic traffic analysis attacks, against mix systems, as inference problems. We have made very good progress so far: we have probabilistic models for the operation of mix networks that incorporate elements of many previously published attacks. Given observations of the network the adversary can use those models to make guesses about who is talking with whom (by applying Bayes theorem.)
A technical report will be ready by December 2009, and then the work will be cut into distinct units for publication.
Metadata for book digitization projects
(Michael Roe)
This is a joint project with the Centre for History and Economics. The overall aim is to find out how a large scale book-scanning project - such as those carried out by Google and Microsoft - should be "done right". From a security perspective, we are interested in what metadata needs to be captured and stored along with digital photographs of pages, in order to answer questions like "What book is this a photograph of?", "Can I rely on it being authentic?", "Who has the rights to duplicate this?"
Michael Roe, Digital Libraries. Proceedings of Security Protocols Workshop 2008.
User-generated content in online games
(Michael Roe)
There is considerable interest in allowing "user generated content" in on-line games. This can range from allowing users to upload bitmap images, all the way to allow end-users to write their own software that hooks into the game and modifies its behavior. The security requirements are similar to those for a distributed multi-user operating system (e.g. want end-users to be able to write reference monitors that mediate access to a piece of stored state), but the setting impose some novel constraints and problems. We are especially interested in the security protocols that are needed when the game is run on multiple servers controlled by organizations that don't trust each other (compare the World Wide Web, which is distributed over a very large number of servers, whose owners don't all trust each other).
