On Sat, Feb 16, 2008 at 01:43:02AM -0500, Roger Dingledine wrote: > > Imagine an adversary who can observe any connection attempt from Alice > and fail any of them that he wants. Imagine this adversary also runs, say, > 10% of the Tor network, including some guard nodes and some exit nodes. > > Alice starts up, learns about the Tor network, picks her entry guards, and > tries to connect to some. Our adversary keeps tricking her into thinking > she picked bad nodes, until she picks an adversary-controlled entry guard. > Then he lets all connections to that entry guard succeed, but when Alice > picks a second hop that isn't adversary-controlled, he claims that next > hop is down. Until eventually he picks an adversary-controlled second > hop. Repeat for the exit node. > > Ignoring bandwidth weightings, exit policies, etc, Alice would need to > try (.1*.1)^-1 = an estimated 100 circuits before she makes three bad > hops (assuming she's already happened across the bad entry guard). For > a more reasonable 1% of the network being bad, that changes to 10000 > circuit rebuilds. > > See http://freehaven.net/anonbib/#ccs07-doa for a related paper here. >
http://freehaven.net/doc/sync-batching/sync-batching.pdf might also give some insight into theses same points, although applied to different network structures, especially in its "Robustness of Anonymity" section, although -Paul
