Re: [Cfrg] J-PAKE and Schnorr NIZK for informational RFCs

[Feng Hao <[e-mail]>]

Hi Watson,

>-----Original Message-----
>From: Watson Ladd [mailto:[e-mail]]
>Sent: 16 November 2016 15:18
>To: Feng Hao <[e-mail]>
>Cc: [e-mail]
>Subject: Re: [Cfrg] J-PAKE and Schnorr NIZK for informational RFCs
>
>On Wed, Nov 16, 2016 at 2:25 AM, Feng Hao <[e-mail]>
>wrote:
>> Hi Watson,
>>
>> On your comments about comparing costs between SPAKE2 and J-PAKE
>>
>>  * It's not a fair comparison as the setup assumptions are different. SPAKE2
>requires a trusted setup, while J-PAKE doesn't. Instead, you should compare
>SPAKE2 with KOY, Jiang-Gong and GL protocols as they require the same
>trusted setup. J-PAKE should be compared with EKE, SPEKE and Dragonfly
>(which is based on SPEKE).
>
>Different setup assumptions do not make it impossible to compare different
>protocols.

OK. I said "fair", not "impossible".

>>  * When you say "SPAKE2 with M and N generated by hashing is secure,
>> and the proofs found in the SPAKE2 paper do work", it's best that you
>write up the full details how you do the hashing, proofs that why it's secure
>and why it doesn't affect the original proofs in the SPAKE2 paper in any way.
>Then people can check and verify your proofs instead of having to take your
>word for it. I see you are always rigorous in asking "security proofs" from
>others for everything they do (which is good), so you should consider
>applying the same rigor to yourself.
>
>This a) cannot be done with standard foundations and b) is a trivial exercise
>with any kind of ideal hash/ROM model. I'll sit down and do it, but while
>drafting this email someone went and did it.

Cool. Btw, nothing is really "trivial" in the design of a protocol. Any small change may have profound effects. This is why I suggest you to write it down in full detail, and ideally get it in a peer-reviewed publication if that's possible so there is a fixed version of the spec as the basis for technical discussion.

>>  * The KOY, Jiang-Gong and GL papers are relevant as they are the same
>type of CRS-based designs as SPAKE2. These papers state that the setup
>needs to be done by a trusted party and they don't specify using a hash.
>> SPAKE2 is in the same model. It's appropriate you compare these protocols
>and definitions. This is necessary especially since you're doing something not
>specified in the original SPAKE2 paper and other related peer-reviewed
>papers.
>
>Except I'm not interested in writing a review paper. I'm interested in looking
>at all the proposals before us for PAKE, and determining which should be
>promoted.
>>
>>  * You will need to convince IETF users that there is no possibility
>> of trapdoor for M and N (which may prove a bit tricky). Knowledge (or
>partial knowledge) of the relation between M and N may allow one to
>systematically break all instances of the protocol execution.
>>
>>  * Kindly note that if you can manage to convince IETF users that M and N
>are completely random, one might just plug M and N as the input of two
>random points to DUAL_EC and make it work?
>
>Are you aware that DUAL_EC included a point generation mechanism that, if
>used, would have completely avoided the backdoor? Do you think that the
>hashing argument presented downthread is wrong.

If that works completely, why not reuse it instead of inventing your own?

>>  * What you say about the 4 exponentiations in the subgroup is correct
>(consistent with the original paper), but this hasn't included the validation
>of the public key (which is free in EC but takes one full exponentiation in the
>finite field setting). See my further comment below.
>
>The next part of my email did describe the validation costs: 1
>exponentiation per element to be validated. That is not a full
>exponentation: if the subgroup is the unique subgroup of order q in the
>finite field, we need only check x^q=1. I said the validation doubled the cost
>of SPAKE2: this is not exactly right, as only one point needs to be validated.

OK. Michel just clarified.

>>
>> I read again the original SPAKE2 paper as well as your I-D. I have the
>> following comments
>>
>>  * I think SPAKE2 is underspecified in the original paper. It doesn't state
>the requirements for M and N, but it should be clear that they must be
>completely random. Also, it doesn't state if the discrete logarithm between
>M and g (and symmetrically between M and g) must be unknown. It's not
>immediately clear to me if knowledge of the discrete logarithm for M and g
>will break anything, but all these should have been explicitly specified in the
>paper.
>>
>>  * I'm a bit worried that in the actual protocol specification in the original
>paper, there is no step to perform public key validation. This could be due to
>two reasons: 1) an inadvertent omission by the authors; 2) an intentional
>design choice. In case of ambiguity like this, one normally takes it as the
>latter. The reason is simple: if public key validation is considered essential, it
>MUST be clearly specified, which is the case with most key agreement
>protocols. However, from early 2000s, some researchers called for
>abandoning the public key validation, as long as the protocol has formal
>security proofs (HMQV is one notable example, but it backfired in the end).
>
>The draft completely specifies an equivalent protection. This is the reason
>we use the cofactor. I believe I explicitly required on-curve checks, but will
>make it more clear if you don't think so.
>
>>
>>  * The above observation reminds me of a paper " Multi-Factor
>Authenticated Key Exchange" by Poincheval and Zimmer in 2008 [1] where
>the protocol is specified in a similar manner as SPAKE2 without public key
>validation. We analysed the protocol and it took us a while to conclude that
>it is insecure without public key validation (despite the security proofs in the
>paper) [2]. We contacted authors of [1] and they kindly acknowledged the
>attack and also confirmed that public key validation needed to be added in
>their protocol. The attack can be traced to a subtle deficiency in their
>theoretical model which implicitly assumes that a server is trusted when it
>communicates with a client. This assumption is clearly invalid, but it is
>implicit in the model/proofs and it took 5 years for peer researchers to
>identify it!
>>
>>  * At the moment, I don't see an obvious attack due to the missing public
>key validation in the SPAKE2 specification (as in the original paper), but it's a
>potential issue that needs some attention.
>>
>> Cheers,
>> Feng
>>
>> [1] D. Pointcheval and S. Zimmer, "Multi-Factor Authenticated Key
>Exchange," Proceedings of Applied Cryptography and Network Security
>(ACNS¹08), pp. 277-295, LNCS 5037, 2008.
>>
>> [2] Security Analysis of a Multi-Factor Authenticated Key Exchange
>> Protocol  https://eprint.iacr.org/2012/039.pdf
>>
>> On 15/11/2016 19:14, "Watson Ladd" <[e-mail]> wrote:
>>
>>>Dear Feng,
>>>
>>>Let me make this very clear, to avoid your misunderstandings: J-PAKE
>>>is substantially less efficient than SPAKE2 over the same group.
>>>SPAKE2 with M and N generated by hashing is secure, and the proofs
>>>found in the SPAKE2 paper do work for this case. If we use a small
>>>subgroup of a finite field group, then the necessary validations for
>>>group membership double the cost of SPAKE2, but J-PAKE is still
>>>slower. J-PAKE requires an additional round, while SPAKE2 fits into
>>>the same flow as Diffie-Hellman. There is no relevance of KOY, or
>>>Jiang-Gong, or any other paper that may or may not (I didn't bother to
>>>look) present its own definitions and security model.
>>>
>>>SPAKE2 requires exactly 4 exponentations in the subgroup if we do not
>>>do anything smart about them. Two of these can be combined and
>>>replaced with a dual base exponentiation via Strauss's algorithm.
>>>
>>>Do you have anything to say to this?
>>>
>>>Sincerely,
>>>Watson
>>
>
>
>
>--
>"Man is born free, but everywhere he is in chains".
>--Rousseau.