Eric,
I thank you for your review. There are some really good insights within
it. I will provide my comments here on how I plan to handle them.
On 11/3/19 3:31 PM, Eric Rescorla wrote:
Sorry for the standalone message. I don't seem to be subscribed to
ietf-announce, so can't reply.
I do not believe that this meets the security standards that we
currently use for designing protocols at IETF. As a general matterE,
this document seems to cut a large number of corners in the service of
some unspecified "overhead" goal, yet it never describes what the
targets are (though presumably this is somehow about computation and
bandwidth), and so it is not possible to evaluate the document against
them. At present, I am unable to say whether it's necessary to do a
new document at all.
The rest of this message details specific deficiencies of this
protocol which should be addressed. Aside from these, it's fairly
unfortunate to see a design for a protocol that uses such an unusual
cryptographic core for no obvious reason. This makes it very hard to
analyze. It would be far better to use SIGMA or some other
well-analyzed construction.
LACK OF PFS
The most serious concern here is the lack of Forward Secrecy. This is
a straightforward static-static DH exchange, but it is not possible to
provide FS in that scenario, as S 1 acknowledges. I have two concerns
here: First, FS is simply table stakes in a modern AKE protocol, so I
don't think we should be publishing a document that doesn't have it
on the standards track.
Second, even if we were to concede that FS might be optional, the
design choices here don't make any sense. There are two major costs to
DH: the cost of doing the DH computations and the bandwidth of sending
the keys. However, given the wide range of the curves permitted here
(X25519 to P521), even a full triple-DH protocol will be far more
efficient on both counts than the existing protocol with P521 (indeed,
it's probably as or more efficient than the existing protocol with
P256).
Proposed action: This protocol needs to be revised to have forward
security.
dex-12.txt will have an applicability statement. You will see text
there that cover work on an 8501 chip with the timings and thus why the
design.
HIT GENERATION
Previous versions of HIP generated HITs from HIs by computing a secure
hash on the HI. This document converts them by a novel folding
procedure. There is no good reason to believe that it is hard to
generate a key that has a given HIT (indeed there are good reasons to
believe it *is* reasonably efficient for non-EC algorithms). The
document sort-of-acknowledges this in Section 9:
o The HIP DEX HIT generation may present new attack opportunities.
Hence, HIP DEX HITs MUST NOT be used as the only means to identify
a peer in an ACL. Instead, the use of the peer's HI is
recommended as explained in Section 3.
However, it's not clear it's sufficient, because nothing strongly
binds the HI (as opposed to the HIT) to the handshake, so attacks
may still be possible if the HI is used (via UKS-like attacks). In
any case, this is a regression from HIPv2.
It's not even clear why this change was made: given that you
have CMAC, you should be able to use this to produce a hash of
the key.
Proposed action: generate HITs from HIs securely.
First point, the HIs are contained in the R1 and I2 packets. Thus they
are bound to the handshake. The HIT is an index to the ACL (or other
lookup mechanism) and duplicates are not allowed. So it is a 'first
come, first allowed' process for HIT to HI mappings. We have always
recognized the risk of duplicates, even with the probability being very
low with a 96 bit truncation of the hash in HIP-BEX.
That said, I considered your recommendation to use CMAC rather than my
novel FOLD operation that I got from another security source. From my
original notes, I remembered that the challenge in using CMAC for HIT
generation is what to use for the key. Should I just use ZERO or some
published random value. So I asked NIST and here is the official response:
"We cannot endorse either choice. The function you describe--CMAC with
a published key--is not an approved cryptographic hash function.
Indeed, It is straightforward to demonstrate that this function is not
collision-resistant for any choice of the key."
So the only reason to switch to CMAC is to satisfy some thought that it
would less likely cause a problem than the FOLD function. Either way
there is need for explaining the need to defend against duplicate HITs,
both 'natural' and forced. And CMAC has a higher computing cost than FOLD.
Right now I feel that I should leave it as is and that the text is
adequate to describe the risk. If you feel I need more warnings, I will
write up text for the security section. If you feel that CMAC should be
used just because, we can discuss it further.
WEAK EC GROUPS
This document specifies the use of SECP160R1. This is not an acceptably
secure group.
Proposed action: REmove support for SECP160R1.
SECP160R1 is carried over from RFC 7401. At the time, there were
participants that wanted something for those really limited CPUs (like
the 8501).
Now we have EC25519 and it can be argued that its performance and
transmitted key size (for constrained networks) meets the needs for
those that wanted SECP160R1. My concern is I have no data on the CPU
cost on a 8501 for SECP160R1 to make a valid comparison to EC25519.
Further the adding of EC25519 does not replace SECP160R1 in HIP-BEX (RFC
7401), so at best, all I could do is not include it in HIP-DEX, it still
would be valid for HIP-BEX.
For now I am leaving SECP160R1 in unless I get feedback that we can
deprecate it for EC25519. Note that a separate draft,
draft-moskowitz-hip-new-crypto, is adding EDDSA25519 for HIP-BEX and
there we can deprecate SECP160R1.
FAILURE TO VALIDATE PUBLIC KEYS
This document does not require that implementations validate
the remote public key. With the NIST curves specified here,
this leads to straightforward key extraction attacks, which
is a very serious problem when you have a static key.
Proposed action: Require point validation. The TLS 1.3 has
text you can borrow.
I am assuming you mean sec 4.2.8.2 of RFC8446?
I am shocked that we have to tell implementers to test that the EC PK is
a valid point. Not.
I will use this text as the basis for an addition subsection in Security
Considerations.
AEAD
This document makes use of non-AEAD symmetric algorithms. This has been
found to be hazardous in practice.
Proposed action: use only AEAD algorithms.
All HIP payloads that occur after there are agreed keys, are
authenticated (MACed). Few payloads have encrypted parameters. Further
the way parameters are organized within the payload it is possible that
the encrypted parameters would not be adjacent, making further
complications. It does not make sense to switch to an AEAD with this
design. Plus an AEAD adds processing overhead. IF NIST had completed
the lightweight crypto competition at design time, then AEAD may have
been given closer study.
REPLAY ATTACK ON AKE
The only entropy provided to the AKE is the puzzle, which means
that it's possible for an attacker to replay the responder's
messages, leaving the initiator believing that he has created
a connection when in fact he has not. The attacker will not
be able to send data messages because the initiator contributes
data to the eventual keys, but we generally try to avoid this
property.
More importantly, this is unnecessary, and can be resolved
by changing the odd "encrypt half the key" mechanism used here
with a conventional nonce structure in which each side sends
a random value and then you HKDF it with the DH shared secret.
This would have the effect of removing the replay attack
and be easier to analyze.
Proposed action: restructure the AKE to mix nonces + DH
into the key schedule.
This is a good suggestion. I am still studying it for the best way to
implement it. I do expect to have something done for a dex-13.txt draft.
-Ekr
Robert Moskowitz
--
last-call mailing list
last-call@xxxxxxxx
https://www.ietf.org/mailman/listinfo/last-call
