From: Ted_Anderson@transarc.com
Date: Tue, 29 Aug 95 10:20:46 PDT
To: dawagner@tucson.princeton.edu
Subject: Re: Cryptanalysis of S-1
In-Reply-To: <41l6u3$852@cnn.Princeton.EDU>
Message-ID: <skEmR=X0BwwMM0o3Im@transarc.com>
MIME-Version: 1.0
Content-Type: text/plain


I find this very interesting.  You have made two related points here
which highlight some important principles of cipher design: (1) more
rounds do not always help and (2) the key schedule can be a limiting
factor in a cipher's strength.  In some sense these are "obvious", but
it helps a lot to have a specific example of these points to think
about. 

After the early looks at S-1 and after reading Blaze & Schneier's paper
on MacGuffin (ftp://research.att.com/dist/mab/mcg.ps) I was thinking
that any half-assed Feistel network could be made secure by adding more
rounds.  So I was thinking about quantifying the systemic cost of adding
more rounds and thereby reducing performance.  It seems that there has
been insufficient analysis of the performance vs. security trade-off. 
In some sense this is understandable given the lack of quantification of
security, but when it comes to engineering a system for real world use,
you have to make a choice and it would be nice to have something to go
on.  Consider for example the use of Blowfish instead of IDEA in
PGPfone; according to Paul Rubin [in "Re: IDEA with PGPFone?",
28-Aug-1995, sci.crypt] this was at least partly due to the performance
difference. 

But here we have a clear limit.  In S-1 the key schedule effectively
limits the number of rounds that contribute to security at about five. 
Further we have a concrete design principle: the per-round sub-keys
should not repeat.  Probably a stronger statement could be made. 

Excerpts from netnews.sci.crypt: 16-Aug-95 Re: S1 cipher P.
Hallam-Baker@w3.org (3569*) 

> I would like to suggest some hypotheses :- 
Maybe this type of cryptanalysis is old hat but it seemed new to me.  It
made me think of another hypothesis for the S-1 release: 
  - It is a training exercise. 
Consider that the primary reason given for keeping Skipjack secret is
that the algorithm would reveal valuable hints about cryptanalysis and
cipher design.  It also seems obvious that the NSA would have a College
of Cyptanalysis to educate new generations of crypto experts.  I could
easily imagine it including a series of exercises, of progressively
increasing difficulty, where attacking each cipher illustrates one or
more cryptographic principles.  Possibly an crypto-anarchist NSA mole
decided it would be safer to leak page from NSA's workbook than Skipjack
itself; an infraction less likely to be persued if nothing else. 

If this seems unlikely, consider that the NSA has been getting beaucoup
bucks for many years now.  With the fall of the "Evil Empire" and all,
perhaps things are getting a bit soft at the core.  Maybe some NSA
strategist figured that a little cross-fertilization between the
academic and national-security crypto communities would enliven both
groups. 

So the question is: Will another exercise appear?  Or perhaps there is
more to learn from this one. 

Ted Anderson