hi,
However, if we add noise that depends on the input, the filtering won't remove it, because for each given input it is constant. Instead of removing the noise component from `code(input) + system_noise()`, leaving `code(input)`, the attacker now deals with `code(input) + hash(input) + system_noise()`, leaving them with `code(input) + hash(input)`. Without cracking the hash, I don't know how you'd remove the hash-based delay from the sample; possible mistakes would be insufficient granularity in the hash-based delay (e.g. if the difference between code paths is 1ms and our delay has a 100ms granularity, it's still trivial to filter out the hash noise), insufficient range (if the difference between code paths is 100ms and our delay is between 0 and 25ms, we still have two distinct groups), a weak hash function, and probably a bunch more.
just a note, that that's what sebastian schinzel's "Deterministic and Unpredictable delay padding" is about. see the talk (i linked to earlier): talk from sebastian schnitzler on 29c3: http://media.ccc.de/browse/congress/2012/29c3-5044-en-time_is_not_on_your_si... slides: http://sebastian-schinzel.de/29c3/download/29c3-schinzel.pdf abstract: http://sebastian-schinzel.de/_download/cosade-2011-extended-abstract.pdf cheers, tob