-
two kinds of side channels:
- (i) software-visible hardware side channels such as caches, pre-fetchers, and branch predictors
- mostly solved by isolating shared micro-architecture
- (ii) software side channels, like page tables == controlled channels
- Xu et al, 2015
- Van Bulck et al, 2017
- idea to solve -> make page tables private (Sanctum, Keystone, ..)
- main claim here -> not enough since OS can still observe sequence of page faults and evictions
-
when do attack conditions occur?
- paging
- disabled in Sanctum so not weak to this attack vector
- at the same time they seem connect it to dynamic memory enclaves???
- applications
have either used demand paging while disregarding the page-
fault side channel [25, 35] or have developed novel oblivious
algorithms that do not leak information via the program’s access pattern [34, 51]
-
tentative solution: InvisiPage -> OPAM (page level ORAM)
-
paper shows that InvisiPage is vulnerable to a new type of side-channel leak.
- It leaks the reuse distances of enclave pages –
the number of page-faults between the moment a page is evicted
and subsequently re-fetched into the enclave.
- they found a pattern in the eviction / paging cycle
- can be exploited in three ways.
- First, we can train a convolutional neural network to fingerprint applications executing within an enclave.
- not needed if you know what's running
- Second, we can use the same kind of fingerprinting to classify the confidential types of inputs an enclave is operating on.
- Finally, we show that the reuse distance attack can be used as covert channel that transmits up to 70 bits/second between an enclave application that is sandboxed using techniques like Ryoan [26] and a colluding operating system.
-
what does keystone do upon page fault?
- trapped at SM level and delegated safely
- so shouldn't be a problem?
-
generic way to mitigate is to put ORAM in middle
