sebastian
commented
4 years ago At the risk of complicating things, I think we might need two delays? Something like this:
total time = (longer of query time and random delay 1) + random delay 2This is in case the difference in the execution time of the two queries is dramatic due to one hitting an optimization of some sort which entirely short circuits the execution and therefore is noticeably (statistically significantly so anyway) different from the other even with the addition of a random delay.
gampleman
yoid2000
One of the bounty teams is trying a timing attack based on small differences in computation (in this case, the difference between the DB returning 1 or 0 rows, and the subsequent processing delay in the cloak). The purpose of this issue is simply to document at a high level what a defense against these kind of "micro" timing attacks. The attack hasn't been run yet so we don't know if it will work.
The reason for documenting a general solution is to prevent someone from patenting a solution and forcing us to pay royalties. (A patent won't apply to us if we already have the solution in-house and can prove when the solution was documented.)
The basic solution idea is to add some delay to answer reporting, where the delay length is determined by the noise layer seeds. The amount of delay needs to be enough to effectively mask whatever delay in computation is being exploited.
Ideally, for micro-timing attacks, the exploited computation difference is quite small and so the added delay could be quite small.
As an example, suppose that the computation delay is 100ms worst case. Then the solution might be to add a constant delay of 200 ms, with a noise value using standard deviation of 100ms and mean zero for each noise layer.
It would be helpful is others of you could add additional ideas just to cover our bases.