[tahoe-dev] [p2p-hackers] convergent encryption reconsidered -- salting and key-strengthening

[zooko]

On Mar 31, 2008, at 4:47 AM, Ivan Krstić wrote:
>
> Tahoe doesn't run this service either. I can't use it to make guesses
> at any of the values you mentioned. I can use it to make guesses at
> whole documents incorporating such values, which is in most cases a
> highly non-trivial distinction.

The way that I would phrase this is that convergent encryption  
exposes whatever data is put into it, in whatever batch-size is put  
into it, to brute-force/dictionary attacks.

If the data that you put in is unguessable, then you needn't worry  
about these attacks.  (Likewise, as Ben Laurie reminds us, using  
strong passwords is a sufficient defense against these attacks on  
passwords.)

You correctly emphasize that typical convergent encryption services  
(which operate on "files", or, in the case of GNUnet, on 32 KiB  
blocks), and typical uses of those services (which typically store  
"files" as produced by apps written for traditional filesystems),  
batch together data in such a way that the aggregate is more likely  
to be unguessable than if each field were stored separately.  I don't  
disagree with this observation.

I am often reminded of Niels Ferguson's and Bruce Schneier's dictum,  
in the excellent _Practical_Cryptography_, that security needs to be  
a *local* property.  They argue that one should be able to tell  
whether a component is secure by inspecting that component itself,  
rather than by reasoning about interactions between that component  
and other components.

Concretely, convergent encryption with a per-user added secret, as  
currently implemented in Tahoe, can be shown to guarantee  
confidentiality of the data, regardless of what the data is.

Traditional convergent encryption can be shown to offer  
confidentiality only with the proviso that the data put into it  
conform to certain criteria -- criteria that cannot be verified by a  
computer nor by a user who is not a skilled security expert.

You may argue that the chance that a user would put non-comformant  
data into it is small.  I don't necessarily disagree, although before  
I became willing to bet on it I would require more quantitative  
investigation.

However, arguing that component A is secure as long as component B  
behaves a certain way, and that component B is very likely to behave  
that way, is a different sort of argument than arguing that component  
A is secure regardless of the behavior of component B.

For one thing, the behavior of component B may change in the future.   
Concretely, people may write apps that store data in Tahoe in a way  
that previous apps didn't.  Those people will almost certainly be  
completely unaware of the nature of convergent encryption and brute- 
force/dictionary attacks.

Now obviously making the security properties of a system modular in  
this way might impose a performance cost.  In the case of Tahoe, that  
cost is the loss of universal convergence.  Allmydata.com analyzed  
the space savings due to convergence among our current customers and  
found that it was around 1% savings.  We (allmydata.com) intend to  
monitor the potential savings of universal convergence in an on-going  
way, and if it turns out that there are substantial benefits to be  
gained then I will revisit this issue and perhaps I will be forced to  
rely on an argument of the other form -- that users are unlikely to  
use it in an unsafe way.

Thank you again for your thoughtful comments on this issue.

Regards,

Zooko O'Whielacronx