[langsec-discuss] Is computation half the story?

travis+ml-langsec at subspacefield.org travis+ml-langsec at subspacefield.org
Mon Mar 30 02:50:14 UTC 2015


On Thu, Mar 26, 2015 at 06:22:59PM -0600, Taylor Hornby wrote:
> ...wherein I make the distinction between a machine's computational
> abilities (i.e. which languages can it decide?) and a machine's
> "informational" abilities (i.e. how can the machine influence the
> outside world? what APIs is it allowed to call?).
> 
> I chose the the term "informational" for lack of a better word because
> it is about information entering and exiting the machine, or moving
> between "parts" of the machine.

I think this is referred to somewhat dismissively as "side effects" of
computation, and it's always been underrated IMHO.

It's actually very, very rare that I am (well, was) actually doing
something that can be modelled as purely functional.  At least half
of the work of any program I worked on was usually I/O of some kind.

Obviously this is not true of biotech informatics and DOE stuff, but
it's definitely true of video games; they are low-latency low-jitter
I/O engines, not offline raytracing engines.

So yes, but so far nobody has really decided what they want to model
and why about I/O.  It's messy and complicated and hard to measure
because there are few guarantees from the underlying hardware.  For
example, on modern hardware, array lookup isn't actually constant
time:

http://cr.yp.to/antiforgery/cachetiming-20050414.pdf

We deal a lot with this in timing side channel attacks:

http://www.subspacefield.org/security/security_concepts/index.html#toc-Subsection-31.2

> I concluded the post by claiming computer science has no general theory*
> of this property. We understand computation well from computability and
> complexity theory, but "informational" capabilities are only understood
> through limited models like ACLs, Bell-LaPaudula, noninterference, etc.
> 
> Those models are properties systems should have in order to be called
> secure. I'm thinking more along the lines of starting with a given
> system then quantifying its its "power" and proving theorems about what
> it can and can't do. Most importantly, relating the power of one given
> system to another given system.

Doing static analysis I deal a lot with "sources" and "sinks" which
deal with what can influence what.  The source is generally some input
to the program, a sink is some sensitive consumer of that data, and
occasionally you detect/model untaint routines that allow it to flow
without problem.  This untainting is rarely as effective as it seems.

I think a common security intuition involves some belief that most
software is already deputized and easily confused, and that the trust
boundaries (i.e. process address space, user/kernel barrier, separate
systems, guest-to-host VM boundary) is our best attempt at "firewalls"
in the original sense of the word.  In addition to static analysis
which models what should happen by spec (code) on an idealized system,
I model what can happen, with the assumption that only systems
explicitly designed to be security barriers tend to be even moderately
effective against knowledgable attackers.  This level of
generalization is necessary for "unknown unknown" deviation from ideal
machines.  It exists in threat modelling and commonly lumped under
"security architecture", but I haven't seen it treated formally.

So in some ways it is more like writing portable code, which deals
with assumptions about the underlying systems and the guarantees they
provide.  That's actually a useful metaphor, because nothing is
universally portable nor universally secure, one has to decide where
to expend effort based on available resources and educated intuition
about potential but hard-to-quantify risks.

> I would appreciate references to the literature.

Can't help you with the theory, as that seems to usually solve
problems I don't have (never seen an infinite tape), but on the
pragmatic side:

"The Art of Software Security Assessment" by Dowd
"Threat Modelling: Designing for Security" by Shostack
-- 
http://www.subspacefield.org/~travis/
"Computer crime, the glamor crime of the 1970s, will become in the
1980s one of the greatest sources of preventable business loss."
John M. Carroll, "Computer Security", first edition cover flap, 1977
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 834 bytes
Desc: not available
URL: <https://mail.langsec.org/pipermail/langsec-discuss/attachments/20150329/56af394f/attachment.sig>


More information about the langsec-discuss mailing list