commit d312aa6609bd5ee3c1026538c6b7fa2dd99e02a5 Author: Mike Perry mikeperry-git@fscked.org Date: Tue Oct 16 13:45:02 2012 -0700

Proposal 209: Fix a math error wrt malicious failure rates.

Forgot I needed to compute failure rates *given* an evil Guard. --- proposals/209-path-bias-tuning.txt | 22 ++++++++++++---------- 1 files changed, 12 insertions(+), 10 deletions(-)

diff --git a/proposals/209-path-bias-tuning.txt b/proposals/209-path-bias-tuning.txt index a1bf493..25ee9de 100644 --- a/proposals/209-path-bias-tuning.txt +++ b/proposals/209-path-bias-tuning.txt @@ -30,13 +30,13 @@ Motivation connections, breaking the O((c/n)^2) property of Tor's original threat model.

- In this case, however, the adversary is only carrying circuits - for which either the entry and exit are compromised, or all - three nodes are compromised. This means that the adversary fails - all but (c/n)^2 + (c/n)^3 of their circuits. For 20% c/n compromise, - such an adversary would only succeed 4.8% of their circuit attempts. - For 33% c/n compromise, such an adversary would still only succeed - 11.7% of their circuits. + In this case, however, the adversary is only carrying circuits for + which either the entry and exit are compromised, or all three nodes are + compromised. This means that the adversary's Guards will fail all but + (c/n) + (c/n)^2 of their circuits for clients that select it. For 10% + c/n compromise, such an adversary succeeds only 11% of their circuits + that start at their compromised Guards. For 20% c/n compromise, such + an adversary would only succeed 24% of their circuit attempts.

It is this property which leads me to believe that a simple local accounting defense is indeed possible and worthwhile. @@ -201,11 +201,13 @@ Security Considerations: Targeted Failure Attacks

Since both conditions would elicit notices and/or warns from all clients, this attack should be detectable. It can also be detected - through the bandwidth authorities, should we deploy #7023. + through the bandwidth authorities (who could possibly even + set pathbias parameters directly based on measured ambient circuit + failure rates), should we deploy #7023.

- We could also conceivably lower pb_disablepct from 30 as a + We could also conceivably lower pb_disablepct to 25% as a potential mitigation, based on the fact that a 20% c/n adversary - would only carry 5% of their circuits in the extreme case. + would only carry 24% of their circuits in the extreme case.

Implementation Notes: Log Messages