Protection for quasi-offline memory nabbing

[Jon Callas]

Such as Cold Boot, etc.

There have been a number of conversations among my colleagues on how  
to ameliorate this, particularly with an eye to making suspend mode  
safer.

In the Cold Boot paper, the authors suggested XORing a piece of random  
memory onto the dangerous bits, so as to fuzz them. This is a clever  
idea, but we didn't like it, particularly because XOR doesn't have the  
best diffusion in the world.

The solution we came up with is to use EME mode (or equivalent) with a  
fixed key. The outline is that you encrypt all dangerous data, like  
volume key, key expansion, etc, with a fixed key into a chunk that you  
keep to the side.

This relies upon the property of EME (and other large-block, tweakable  
modes) that a single-bit error in the ciphertext propagates to an  
error in the entire plaintext. Consequently, a very low rate of memory  
decay turns into complete protection of that sensitive data.

Upon suspend, you erase and deallocate the active store, and on wake  
you decrypt the fuzzed copy to get your keys and state variables back.

If you want to one-plus this, you could have a timeout on the drive so  
that if it's idle for N seconds, you do same.

When we came up with this, we wondered if it was patentable. We've  
decided that it isn't, that this is something that is obvious to  
someone skilled in the art. Our reasoning is something like:

Cold Boot paper suggests XORing random memory
	but -- XOR has cruddy diffusion
What has better diffusion?
(discard suggestions like lead, churches, and very small rocks)
Block ciphers have great diffusion
	but -- block ciphers operate on only a small chunk
What operates like a block cipher on a large chunk?
Tweakable modes like EME.
QED

The rest is just software engineering.

The cool thing about using EME (or equivalent) is that the larger the  
chunk you create, the better you survive a Cold Boot attack.

Note, however, that an attacker who can grab memory with no errors in  
it, such as someone who is playing DMA games, still gets the keys. To  
protect against that, you have to have an authentication mechanism,  
which is outside the scope of this -- we want something that is  
transparent, but can make people worry less about suspending their  
laptop.

Also note that you don't really need a full cipher. All you need is  
reversible diffusion that maximizes damage on a single-bit error.  
However, the danger in coming up with another function is that you're  
effectively designing special-purpose crypto. Yes, it's really special- 
purpose coding, not crypto, but it's a lot safer to use crypto. We  
understand it better.

A number of people participated in our discussions and at least two  
people independently thought of the core idea. The people include but  
are not limited to (which means I apologize to everyone I forgot):  
Colin Plumb, Phil Zimmermann, Hal Finney, Andrey Jivsov, Will Price,  
David Finkelstein, and Bill Zhao.

	Jon

---------------------------------------------------------------------
The Cryptography Mailing List
Unsubscribe by sending "unsubscribe cryptography" to majordomo at metzdowd.com