[Bitcoin-development] BIP 32.5

[Gregory Maxwell <gmaxwell@gmail.com>]

I am wondering if we shouldn't have a BIP32 addendum which makes the
following signing related recommendations:

(1) Recommend a specific deterministic DSA derandomization procedure
(a deterministic way to generate the DSA nonce), presumably one based
on HMAC-SHA512 (since BIP32 uses that construct) or SHA256 in the
style of RFC 6979.

DSA systems being compromised due to poor randomness at runtime is not
new. It effected other systems before it effected Bitcoin systems,
it's not a new problem and it's not going away.  It's difficult to
tell if an implementation is correct or not.

Use of a fully deterministic signature  would allow for complete test
vectors in signing and complete confidence that there is no random
number related weakness in a signing implementation.

In particular, with relevance to our ecosystem a maliciously modified
difficult to audit hardware wallet could be leaking its keys material
via its signatures. Even without producing insecure K values it could
use the choice of K to leak a couple bits of an encrypted root key
with every signature, and allow the malicious party to recover the
keys by simply observing the network. Making the signatures
deterministic would make this kind of misbehavior practically
discoverable.

We wouldn't be alone in making this change, in general industry is
moving in this direction because it has become clear that DSA is a
hazard otherwise.

The primary arguments in most spaces against derandomizing DSA are
FIPS conformance (irrelevant for us) and reasonable concerns about the
risks of using a (less) reviewed cryptographic construct. With
widespread motion towards derandomized DSA this latter concern is less
of an issue.

Libcrypt has also implemented derandomized DSA in git. The ed25519
signature system of DJB, et. al. also uses a similar derandomization.

An alternative is implementing a still random construct where K is
some H(message||key||random) which should remain secure even where the
randomness is poor, but this loses the advantage of being able to
externally verify that an implementation is not leaking information.
OpenSSL development has implemented a form of this recently.

See also: http://tools.ietf.org/rfc/rfc6979.txt

(2) Recommends a procedure for using only even S values in signatures,
eliminating this source of mutability in transactions.

This can be accomplished via post-processing of existing signatures,
but since it requires bignum math it is usually preferable to
implement it along with signing.  I believe someday this will become a
network requirement for Bitcoin, but regardless it makes sense to
implement it as a best practice sooner rather than later.

Thoughts?