Round 1 (Additional Signatures) OFFICIAL COMMENT: EagleSign

3,653 views
Skip to first unread message

Mehdi Tibouchi

unread,
Jul 17, 2023, 12:32:08 PM7/17/23
to pqc-co...@nist.gov, pqc-...@list.nist.gov
Dear EagleSign submitters, dear all,

The EagleSign signature is a Fiat-Shamir construction without aborts that
uses small noise and a polynomially-bounded modulus; as such, it is not
plausibly secure.

Leakage of the secret key in signatures occurs for example in
signature element z, which is of the form:

z = G·(y_1 + c)

where G is a short matrix part of the secret key (and recovering it is
effectively sufficient for full key recovery), c is known and small, and
the y_1 is small with known, independent distribution. No modular
reduction occurs in this equation.

As a result, the distribution of z depends on G in a way that easily
reveals it using statistical techniques (for example, collect many
signatures for which the first coefficient of c is 1, which is satisfied
by a significant fraction of all signatures, and observe that the
conditional expectation depends linearly on G). The same attack can be
mounted on the other signature element w.

Accordingly, the security proof is incorrect. For example, the element z
of the signature is simulated as a uniformly random among the elements of
R of infinity norm up to the maximum possible according to the equation
above, but clearly, the real distribution is not uniform at all.

Best regards,

--
M. Tibouchi
<mehdi.t...@normalesup.org>
http://www.normalesup.org/~tibouchi/
************************************************************
NTT Social Informatics Laboratories
Abe Research Laboratory
3-9-11 Midori-cho, Musashino-shi, Tokyo, 180-8585, Japan.
************************************************************

Mehdi Tibouchi

unread,
Jul 18, 2023, 12:03:51 AM7/18/23
to pqc-co...@nist.gov, pqc-...@list.nist.gov
Dear eMLE-Sig 2.0 submitters, dear all,

The eMLE-Sig scheme is a lattice-based scheme that seems to be a
spiritual successor to the original Round 1 encryption candidate Compact
LWE. The design principle of this type of schemes is essentially to say:
“we will use as a trapdoor a lattice vector that isn't particularly
short, so that there is no hope to recover it by lattice reduction, but
that has some sort of additional hidden structure that still allows it to
be used as a secret; moreover, since lattice attacks must fail, we will
use a ridiculously small lattice dimension”.

The underlying belief is that the additional hidden structure cannot be
taken advantage of in lattice attacks. This belief, however, proved
incorrect in the case of Compact LWE (which co-authors and I broke
shortly after submission), and again in the first instance of eMLE which
was broken on this mailing list by Lorenz Panny a couple of years ago.

It is incorrect again in eMLE-Sig 2.0. Basically, a secret key element x
in that scheme is a short element of the “convolution” ring Z[x]/(x^n-1)
satisfying an equation of the form:

h' = g'·x + p_0·k'_0 + p_1·k'_1 + p_2·k'_2 (*)

where g' and h' are known ring elements, the p_i's are rational primes
(with p_2 >> p_1 >> p_0) and the k'_i's are unknown but short.

The break on the original version of the scheme was based on the
observation that the equation (*) above can be seen as an
Inhomogeneous-SIS instance mod p_2 recovering the short vector
(x,k'_0,k'_1). This attack is expected to succeed when the length of that
vector is substantially smaller than the first minimum of the SIS
lattice, which by the Gaussian heuristic is approximately:

lambda_1(L_{SIS}) = sqrt{n/2\pi e} p_2^{1/3}.

The countermeasure proposed in the eMLE-Sig 2.0 submission is to add
noise making k'_1 larger, so that (x,k'_0,k'_1) is no longer below
lambda_1 above. Concretely speaking, the “level-I” parameter have roughly
the following vector Euclidean norms:

|x| \approx 20
|k'_0| \approx |k'_2| \approx 45
|k'_1| \approx 1300
lambda_1 \approx 800

So indeed, (x,k'_0,k'_1) is of length larger than lambda_1, and one
expects exponentially many vectors of that length in the SIS lattice.
However, it is a very structured in the sense that x and k'_0 are
themselves much shorter, and this can obviously be taken advantage of.
The highbrow way to do this is to scale the Euclidean norm itself and
adjust the volume computations, etc., accordingly. A similar, more
pedestrian way is to scale the subvectors by integer factors: if a, b are
coprime to p_2, (a·x, b·k'_0, k'_1) is a solution to the ISIS problem:

h' = a^{-1} g'·u + b^{-1} p_0·v + p_1·k'_1 (mod p_2)

that moreover lies in aZ^n × bZ^n × Z^n. The normalized volume of the
corresponding lattice increases from p_2^{1/3} to (ab p_2)^{1/3}, and
hence the lambda_1 is multiplied by (ab)^{1/3} compared to the above.
Picking a = 65, b = 29 to balance out all the subparts of the vector, we
end up looking for a vector of norm \sqrt{3}·1300 ~ 2250 in a lattice
with

lambda_1 \approx 800·(ab)^{1/3} \approx 10000

which is now a well-defined problem with a unique solution whp. Given the
very small lattice dimension 3n = 192, it is also tractable in practice.

I note in addition that, even if the eMLE problem were hard, the proposed
signature scheme could not be, since, just like in my previous email
regarding EagleSign, the distribution of signatures leaks information
about the secret key.

Mehdi Tibouchi

unread,
Jul 19, 2023, 2:23:10 AM7/19/23
to pqc-co...@nist.gov, pqc-...@list.nist.gov
Dear EagleSign submitters, dear all,

For the sake of completeness, you can find sample code that implements
the attack below in the following GitHub repository:

https://github.com/mti/attack_eaglesign

As mentioned in the README, this is a fairly naive implementation of the
idea, but it already correctly recovers around 1020 coefficients of G out
of 1024 for parameter set EagleSign-3 with 100,000 signature samples on
arbitrary messages.

Best regards,

--
M. Tibouchi
<mehdi.t...@normalesup.org>
http://www.normalesup.org/~tibouchi/
************************************************************
NTT Social Informatics Laboratories
Abe Research Laboratory
3-9-11 Midori-cho, Musashino-shi, Tokyo, 180-8585, Japan.
************************************************************


Ludo Pulles

unread,
Jul 19, 2023, 5:19:54 AM7/19/23
to Mehdi Tibouchi, pqc-co...@nist.gov, pqc-...@list.nist.gov
Hi Mehdi,

Great work finding this. I think the outlined attack is already quite disastrous. Just to remove any doubt whether recovering all but a couple of the coefficients really breaks the scheme:

https://github.com/ludopulles/EagleHasFlown

Here is an implementation that throws away less of the signatures (re: your 2nd remark in the README): you can use the same statistical argument for all the other coefficients of c as well, greatly reducing the number of (random) signatures required for the attack to work.
The matrix G is recovered completely with 300 and 500 signatures for EagleSign3 and EagleSign5 respectively, and the matrix D is fully recovered with 2000 signatures for both EagleSign-{3,5}.

Best regards,
Ludo
 

Van: Mehdi <mehdi.t...@normalesup.org>
naar: pqc-comments <pqc-co...@nist.gov>
CC: pqc-forum <pqc-...@list.nist.gov>
datum: woensdag 19 juli 2023 8:23 CEST
Onderwerp: [pqc-forum] Re: Round 1 (Additional Signatures) OFFICIAL COMMENT: EagleSign
--
You received this message because you are subscribed to the Google Groups "pqc-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pqc-forum+...@list.nist.gov.
To view this discussion on the web visit https://groups.google.com/a/list.nist.gov/d/msgid/pqc-forum/ZLeBSPnwt9lJ6y9l%40phare.normalesup.org.

djiby sow

unread,
Jul 19, 2023, 1:55:28 PM7/19/23
to Mehdi Tibouchi, pqc-co...@nist.gov, pqc-...@list.nist.gov
Hi all 
Great work Tibouchi for this attack on EagleSign and for your Quick implementation of this attack . Impressive.  
Since EagleSign involve other variants, we are testing the zero knowledge property for these variants 
Best regards .

djiby sow

unread,
Jul 19, 2023, 2:04:13 PM7/19/23
to Ludo Pulles, Mehdi Tibouchi, pqc-co...@nist.gov, pqc-...@list.nist.gov
Hi all 
Great work Pulles for your work by improving the implementation of the attack of Tibouchi on EagleSign.  
As said before we are working  other variants that hold the zero knowledge property 
Best regards

Mehdi Tibouchi

unread,
Jul 19, 2023, 11:50:47 PM7/19/23
to pqc-co...@nist.gov, pqc-...@list.nist.gov
Dear eMLE-Sig 2.0 submitters, dear all,

For the sake of completeness again, you can find sample code that
implements the attack below in the following GitHub repository:

https://github.com/mti/attack_emle

For now, it targets the "NIST Level-I" parameter set, and recovers a
signing key from the corresponding verification key with fairly good
probability (>80%) in a few minutes. Larger parameters should be broken
by essentially the same attack code, but this hasn't been verified yet.

Best regards,

--
M. Tibouchi
<mehdi.t...@normalesup.org>
http://www.normalesup.org/~tibouchi/
************************************************************
NTT Social Informatics Laboratories
Abe Research Laboratory
3-9-11 Midori-cho, Musashino-shi, Tokyo, 180-8585, Japan.
************************************************************

Liu, Dongxi (Data61, Marsfield)

unread,
Jul 20, 2023, 1:41:56 AM7/20/23
to Mehdi Tibouchi, pqc-co...@nist.gov, pqc-...@list.nist.gov

Hi  M.  Tibouchi,

 

Thanks for your analysis. 

 

Simply, your attack is NOT correct (at least for your current version), because you set q=256. 

Note that q is just used for security analysis in Sage code and it is not a parameter in the specification and in the c-language reference implementation. In Sage code,  we test different values of q to do security analysis and determine the concrete security level. A smaller q means smaller amount of noise than designed is added.  The hardness of eMLE 2.0 depends on both dimension and the amount of noises, a difference from lattice problem. 

 

As mentioned in the specification, when q=256, the attack code in our submitted package (adapted from Panny’s attack) is 100%, better than your 80%.  Different values of q is discussed in Section 5.3.1 of the specification. Our attack code  actually exhaustively search your “a”; so I think your attack has been covered in our attack analysis and security estimation.  

 

Also note that in Table 2 of our specification , the norm of k1 is usually above 11000 for n=64; not  “>   |k'_1| \approx 1300” as you said in your first email. 

 

If we use q=2^20 to allow all noises as designed in the specification, the result of your attack on my machine is like this, 0% success rate, run twice:

 

Attack against eMLE-Sig 2.0 keys

--------------------------------

Precompute key-independent reduction:

| ...done

Instance 1/25:

| x1 = (0, 4, 3, -4, 0, -1, 0, 2, 2, 4, 4, -3, -3, -4, -4, ...)

| x2 = (2, 0, -1, 3, -1, 0, 0, 2, 3, -3, 0, -2, 4, 0, -2, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 2/25:

| x1 = (4, 0, -1, 4, 4, 2, -4, -1, 3, -2, -4, -1, -4, 4, 1, ...)

| x2 = (2, 4, 2, -4, 4, -2, -3, -4, -3, 1, -1, 2, 3, -4, -1, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 3/25:

| x1 = (-2, -2, 0, -3, -4, -1, 1, -1, -1, -1, -4, -4, -4, -4, -1, ...)

| x2 = (2, -2, -3, 0, -2, 3, -3, 4, 3, 1, 3, -3, -2, 1, 3, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 4/25:

| x1 = (1, -3, 2, 4, -2, 0, 1, 3, 2, -4, -2, 1, 4, 3, -2, ...)

| x2 = (0, -2, 1, -4, -3, 0, -3, 2, -3, -2, 0, 1, -4, 3, -3, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 5/25:

| x1 = (4, -2, 4, -1, -2, -3, -3, 2, -1, 0, 3, -2, 0, 0, 3, ...)

| x2 = (0, -2, 4, -2, -2, -3, -2, 1, -2, -3, -3, -2, 0, 3, 1, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 6/25:

| x1 = (2, -3, 0, 4, 1, -4, 1, 3, 2, 3, -4, 0, -1, -1, 3, ...)

| x2 = (-3, 0, 4, -4, -4, -1, -1, 3, 2, 0, 4, -4, -4, -4, 0, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 7/25:

| x1 = (-3, -2, 1, -2, 0, 0, 2, 1, 4, 2, -4, -2, -2, 2, 2, ...)

| x2 = (2, -1, 4, -1, -1, -4, 3, 3, -4, 0, 4, 3, -1, -1, -4, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 8/25:

| x1 = (-4, 1, -3, 4, 3, 0, -4, 3, 1, 1, 0, 2, -2, -1, 3, ...)

| x2 = (4, 2, -4, 3, 4, 4, 2, -3, 2, -2, 1, -1, -1, 0, -2, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 9/25:

| x1 = (-2, -1, 2, -2, -1, 1, 2, -1, -2, -2, 0, 4, 4, 1, 4, ...)

| x2 = (4, -1, -4, -4, -2, -3, 3, 1, -1, -3, 4, 3, -4, -2, -4, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 10/25:

| x1 = (2, -4, 0, 0, 1, 3, 3, 1, 4, -4, 3, 2, 2, 1, -3, ...)

| x2 = (-1, -4, -1, 3, -2, -4, -3, 4, -1, -3, 2, -2, -4, 3, 1, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 11/25:

| x1 = (3, -4, 4, -4, -1, 1, 3, 0, 4, -4, -4, 3, -1, 2, -4, ...)

| x2 = (0, -4, -2, -4, 1, 2, -1, -3, -4, 3, 3, 0, 3, -3, 4, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 12/25:

| x1 = (-3, 1, 4, 2, 4, -4, 1, -4, -2, 4, 0, -2, -1, -3, 1, ...)

| x2 = (-1, 4, -1, 0, 3, -4, -2, -4, 3, -2, -3, 1, -1, 3, 0, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 13/25:

| x1 = (-4, -4, 2, -3, -1, 1, 0, -1, 3, 1, -3, -3, -2, -3, -4, ...)

| x2 = (1, 4, 1, 4, -2, -3, -2, -3, -3, -1, 2, 3, -1, 3, -1, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 14/25:

| x1 = (4, 1, -1, 1, 3, 1, 2, 4, -2, -4, -2, -1, 4, -1, 4, ...)

| x2 = (1, -3, -1, -3, 1, -4, 2, 3, -2, -4, -2, -1, 0, -2, -1, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 15/25:

| x1 = (3, 3, -2, 4, 2, -1, 3, 2, -1, -4, 0, -3, -2, 0, 4, ...)

| x2 = (-1, 0, -1, 4, -1, 3, -4, 3, -4, -3, 1, 2, -2, -1, -4, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 16/25:

| x1 = (-4, 4, 2, 3, -3, 1, -1, -3, 4, -2, -2, -1, 4, 3, -4, ...)

| x2 = (1, 0, 1, -4, 4, 2, -1, -2, -1, 2, 3, -1, -4, 3, 2, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 17/25:

| x1 = (-3, 0, 1, 2, 2, -2, 1, -1, 1, -1, -1, 4, -1, -4, -2, ...)

| x2 = (-4, 4, 2, 1, -4, -1, -4, 0, -1, 1, -4, 4, -1, 2, 4, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 18/25:

| x1 = (-2, 1, -1, -4, 3, -4, -2, 3, 3, 1, -4, -2, 0, 2, 2, ...)

| x2 = (-4, 3, -3, 3, 0, 4, 3, 3, -4, 2, 4, 0, -3, 4, -3, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 19/25:

| x1 = (-2, 4, -3, -1, 2, -1, -4, 1, 2, 1, 4, 4, 2, -3, -4, ...)

| x2 = (0, -1, -2, 0, -1, 0, 4, -4, 1, 1, 2, -2, -3, 3, 0, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 20/25:

| x1 = (0, 4, 0, -3, -3, 2, 2, -1, -3, -2, -3, -4, 2, 4, 0, ...)

| x2 = (-2, -4, 3, 2, 2, -4, -4, -1, 3, 3, 3, 2, -1, 2, -4, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 21/25:

| x1 = (4, 2, -3, 3, -2, 2, -3, 2, -4, -1, 4, 3, 1, 0, 2, ...)

| x2 = (4, -2, 0, 1, -2, 2, -1, 3, 3, 0, 3, 0, -4, 4, 2, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 22/25:

| x1 = (-1, 1, -1, 4, -3, -1, -2, -3, 3, -2, 3, 2, 2, -3, 0, ...)

| x2 = (-1, -1, 3, -2, 1, 2, -2, -3, 0, -3, -3, 2, -1, 2, -4, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 23/25:

| x1 = (-3, 1, -4, -3, 0, -3, 3, 3, -2, 1, 4, 3, 4, 4, 2, ...)

| x2 = (2, 0, -4, -4, -4, 0, 3, 1, -1, -3, -2, 4, -3, -4, 2, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 24/25:

| x1 = (2, -4, 4, -3, -4, -1, 2, -2, 3, 0, 3, 1, -2, 0, -4, ...)

| x2 = (-1, 4, -3, 3, 4, -4, 3, -3, 4, -1, 1, 1, 4, -2, -1, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

Instance 25/25:

| x1 = (3, 3, -1, -3, -3, 3, -2, 4, 2, 4, -1, -1, -3, 1, 1, ...)

| x2 = (4, -2, 4, 4, -1, 1, -2, 2, 1, 1, 3, 2, -1, -3, -4, ...)

K0:  6975 5096

| x1 recovery failed

K0:  6975 5096

| x2 recovery failed

0/50 correct recoveries (0.0% success rate)

 

Regards,

Dongxi Liu

 

 

--
You received this message because you are subscribed to the Google Groups "pqc-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pqc-forum+...@list.nist.gov.

Mehdi Tibouchi

unread,
Jul 20, 2023, 6:22:14 AM7/20/23
to Liu, Dongxi (Data61, Marsfield), pqc-co...@nist.gov, pqc-...@list.nist.gov
Dear eMLE-Sig 2.0 submitters, dear all,

I acknowledge that I had misunderstood the provided SageMath code, and
the attack does not work as proposed. My apologies.

It might in fact be the case that the eMLE problem (in the sense of
recovering x from h) is hard. If so, however, I am fairly convinced that
you will not be able to build actual cryptographic schemes that are as
hard to break as the problem itself.

In the case of eMLE-Sig, as I mentioned earlier, the scheme suffers from
clear secret key leakage in signatures. Example code demonstrating this
has now replaced the previous, incorrect attack on the following
repository:

https://github.com/mti/attack_emle

It basically recovers the full signing key in the "NIST Level-I"
parameter set with around 2,500,000 signature samples.

Best regards,

--
M. Tibouchi
<mehdi.t...@normalesup.org>
http://www.normalesup.org/~tibouchi/
************************************************************
NTT Social Informatics Laboratories
Abe Research Laboratory
3-9-11 Midori-cho, Musashino-shi, Tokyo, 180-8585, Japan.
************************************************************

Liu, Dongxi (Data61, Marsfield)

unread,
Jul 21, 2023, 3:12:30 AM7/21/23
to Mehdi Tibouchi, pqc-co...@nist.gov, pqc-...@list.nist.gov

Hi M. Tibouchi,

 

Your leakage attack works.  Thanks for further analysis. 

 

Regards,

Dongxi Liu

Lorenz Panny

unread,
Jul 23, 2023, 4:25:45 AM7/23/23
to pqc-co...@nist.gov, pqc-...@list.nist.gov
Dear all,

here's software that computes eMLE-Sig 2.0 secret keys from public
keys within a few (single-core) hours per keypair at level 1:

https://yx7.cc/files/emle2-attack.tar.gz

The software implements the attack outlined by Mehdi Tibouchi in his
comment from 2023-07-18 on this mailing list:

https://groups.google.com/a/list.nist.gov/g/pqc-forum/c/zas5PLiBe6A/m/EVmNzzglBQAJ

To find the short target vector, the code runs progressive BKZ tours
up to block size 90 using the G6K library. On average, a block size
around 80 seems to suffice.

I've tested the software against the 100 public keys from the KATs.
It successfully recovered 83 secret keys, which have been confirmed
capable of forging signatures, and failed in 17 cases for various
non-fundamental reasons (higher precision required for fplll not to
crash, or stronger reduction required to find the target vector).

To run the attack, a working setup of SageMath version 10.0 with the
G6K library is required. For convenience, a Dockerfile to run such
an environment is included. Thus, a given level-1 public key can be
attacked by simply running

./run.sh A821C296...BF8A0ABB

where A821C296...BF8A0ABB is the public key (given in the format of
the KATs). After it has finished, the resulting secret key can be
tested for validity by running

./check.py A821C296...BF8A0ABB 000201FC...2760238D

where 000201FC...2760238D is the secret key found by the attack.

Best,
Lorenz

Liu, Dongxi (Data61, Marsfield)

unread,
Aug 1, 2023, 12:48:51 AM8/1/23
to Lorenz Panny, pqc-co...@nist.gov, pqc-...@list.nist.gov

Hi Lorenz, Hi M. Tibouchi, 

 

Thanks for your attacks.

eMLE-Sig 2.0 has been revised to address the attacks. 

 

The major revision is summarized below. Welcome further analysis (from anyone).

 

======================================================

1. Leakage attack 

 

[current]   s = x1*c1 + x2*c2 + y

[revised]   s = x1*c1 + x2*c2 + (c1+c2)*y

 

The leakage attack now finds x1-x2 (checked with Tibouchi’s code), no longer individual x1 and x2. So one of x1 or x2 has to be guessed for being able to forge signatures. Together with the revision below, there are about 380 bits to guess for level 1.

 

2. Lattice reduction attack 

 

The revision is to aggregate public keys and share a new secret z across two eMLE values in a public key.     

 

—Private key

[current]  x1, x2

[revised] (x1, x1’),  (x2, x2’), z

 

—Public key

[current]  h1 = eMLE(G, x1, G[1]), h2 = eMLE(G, x2, G[1])

[revised]  h1 = eMLE(G, x1, G[1]+z)+ eMLE(G’, x1’, 0),  h2 = eMLE(G, x2, G[1]+z) + eMLE(G’, x2’, 0) 

 

G[1]+z is embedded into all layers. Elements of z are from 0 to p[2]-1. 

The shared z can be removed by h1-h2. But by this, at the best case the attack recovers x1-x2, and x1’-x2’ (the same result as leakage attack). One of (x1, x1’) or (x2, x2’) needs to be guessed.  

 

—Signing 

 [current] (u, s),      where u = eMLE(G, y, c1’+c2’),   s = x1*c1 + x2*c2 + y

 [revised] (u, s, s’), where u = eMLE(G, y, c1’+c2’-z) + eMLE(G’, y’, 0),   

                                        s = x1*c1 + x2*c2 + (c1+c2)*y,

                                        s’ = x1’*c1 + x2’*c2 + (c1+c2)*y’

 

 

3. Sizes of public key and signature for the revised version 

 

p[2] changes from 2^26 to 2^22.  New sizes are listed below (bigger signature, but smaller pk).

 

[level 1]  pk:  2*22 * 64/8  = 352 bytes,       sig:  (22 + 2*9)*64/8 = 320 bytes

[level 3]  pk: 2*24 * 96/8  = 576 bytes,        sig:  (24 + 2*10)*96/8 = 528 bytes

[level 5]  pk: 2*26 * 128/8 = 832 bytes,       sig:  (26 + 2*10)*128/8 = 736 bytes

 

 

The revised sage implementation is available at:  

https://gitlab.com/raykzhao/emle-sig/-/blob/main/Extra/sage-code/impl-gen.sage

https://gitlab.com/raykzhao/emle-sig/-/blob/main/Extra/sage-code/correctness.sage 

 

======================================================

 

Regards

Dongxi  Liu

 

From: 'Lorenz Panny' via pqc-forum <pqc-...@list.nist.gov>


Date: Sunday, 23 July 2023 at 6:25 pm
To: pqc-co...@nist.gov <pqc-co...@nist.gov>
Cc: pqc-...@list.nist.gov <pqc-...@list.nist.gov>

--
You received this message because you are subscribed to the Google Groups "pqc-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pqc-forum+...@list.nist.gov.

Mehdi Tibouchi

unread,
Aug 2, 2023, 2:12:45 AM8/2/23
to Liu, Dongxi (Data61, Marsfield), Lorenz Panny, pqc-co...@nist.gov, pqc-...@list.nist.gov
Dear eMLE-Sig 2.0 submitters, dear all,

I find the description of the revision below a bit difficult to follow,
so pardon me if I misunderstand, but I fail to see how the changes
address the leakage attack at all.

I hadn't explained the attack in much details before, so let me describe
it on the original scheme. Signatures in the original submission contain
an element of the form:

s = x_1·c_1 + x_2·c_2 + y,

where c_1 and c_2 are known outputs of a random oracle, y is a random
mask (which is unknown), and x_1, x_2 are the secret key elements. All
variables are elements of the “convolution” ring R = Z[x]/(x^n-1).

The coefficients of c_1 and c_2 (in the polynomial basis of R) are
uniformly and independently distributed in {0, 1, 2, 3}. The coefficients
of y have a somewhat more complicated distribution, but they are i.i.d.
(and independent of c_1, c_2 thanks to the random oracle).

By the law of large numbers, if we generate many signatures and average
everything out, the resulting value will converge towards the
expectation:

E[s] = x_1·E[c_1] + x_2·E[c_2] + E[y].

In this equation, the expectations of E[c_1] and E[c_2] are the same and
easy to compute: they're the ring element with all coefficients equal to
the mean of the uniform distribution on {0, 1, 2, 3}, namely (3/2)w where
w = (1,1,...,1) = 1 + x + ··· + x^{n-1}. Similarly, since the
coefficients of y are i.i.d., E[y] is some multiple mu·w for some real
number mu (which we can estimate, but it doesn't matter).

If instead I restrict attention to signatures such that the first (i.e.,
constant) coefficient of c_1 is, say, 3, everything remains the same,
except that E[c_1] in now (3, 3/2, 3/2, ..., 3/2) = 3/2(1+w). Thus,
averaging many such values of s yields a value close to:

e_3 = E[s | c_{1,0} = 3] = (3/2) x_1 + w·((3/2)x_1 + (3/2)x_2 + mu).

Now, w is a zero divisor in R, with for example z·w = 0 for z = 1 - x =
(1,-1,0,...,0). Thus, we can compute:

z·e_3 = (3/2) z·x_1 + z·w·((3/2)x_1 + (3/2)x_2 + mu) = (3/2) z·x_1,

and from that, recovering x_1 is easy since it suffices to guess a
single coefficient in {-4,...,4} to recover the whole vector.

Ok, so now, how does the proposed change, namely:

> s = x1*c1 + x2*c2 + (c1+c2)*y,

affect this attack? Not much as far as I can tell from the description
alone. By independence, we now have:

E[s] = x_1·E[c_1] + x_2·E[c_2] + (E[c_1] + E[c_2]) · E[y],

and similarly for conditional expectations, so the same trick should
reveal x_1, and not just x_1-x_2 as claimed below. Therefore, I am
curious about the following statement:

> The leakage attack now finds x1-x2 (checked with Tibouchi’s code), no
> longer individual x1 and x2.

Please share the code used to arrive at this conclusion. (It is not
practical to test anything, especially not statistical attacks requiring
millions of signatures, based on the toy SageMath implementation alone,
which is slow and clunky).


Regarding the lattice attack, an updated version of the specification
would be appreciated, as well as a clarification of the security claims.
To be clear, are you still claiming hardness for some underlying "eMLE
problem", or are you just saying that recovering the three unknowns x1,
x1' and z from the sole datum of h1 is difficult? (Depending on the
parameters, the latter might conceivably be information theoretically
hard, but in that case, again, there is no hope of building cryptographic
schemes from it).

Best regards,

--
M. Tibouchi
<mehdi.t...@normalesup.org>
http://www.normalesup.org/~tibouchi/
************************************************************
NTT Social Informatics Laboratories
Abe Research Laboratory
3-9-11 Midori-cho, Musashino-shi, Tokyo, 180-8585, Japan.
************************************************************


> To view this discussion on the web visit https://groups.google.com/a/list.nist.gov/d/msgid/pqc-forum/ME3PR01MB6919D16C99D3E1F42206EF37D00AA%40ME3PR01MB6919.ausprd01.prod.outlook.com.

Liu, Dongxi (Data61, Marsfield)

unread,
Aug 2, 2023, 4:56:32 PM8/2/23
to Mehdi Tibouchi, Lorenz Panny, pqc-co...@nist.gov, pqc-...@list.nist.gov

Hi M. Tibouchi,

 

> Please share the code used to arrive at this conclusion.

 

Here are the files:

https://gitlab.com/raykzhao/emle-sig/-/blob/main/Extra/sage-code/leakage/impl.c

https://gitlab.com/raykzhao/emle-sig/-/blob/main/Extra/sage-code/leakage/test_attack.c

 

[impl.c] the sign function is revised to implement “x1*c1 + x2*c2 + (c1+c2)*y”, with y’s elements limited from -4 to 4.

 

[test_attack.c] your file, with modification to recover also x2zrec, and then print x1z, x2z, x1zrec, x2zrec, x1z-x2z, x1zrec – x2zrec.

 

You just put these two files into your package; all other steps are the same. 

Thanks for your checking. I will answer your other questions after this.

 

 

Regards,

Dongxi

 

 

From: Mehdi Tibouchi <mehdi.t...@normalesup.org>


Date: Wednesday, 2 August 2023 at 4:12 pm
To: Liu, Dongxi (Data61, Marsfield) <Dongx...@data61.csiro.au>

Mehdi Tibouchi

unread,
Aug 2, 2023, 10:08:08 PM8/2/23
to Liu, Dongxi (Data61, Marsfield), Lorenz Panny, pqc-co...@nist.gov, pqc-...@list.nist.gov
Dear Dr. Liu,

Thanks for the modified implementation. A corresponding attack has been
pushed to the following branch of my repository:

https://github.com/mti/attack_emle/tree/revisedscheme20230822

As you can see, and in accordance with the analysis I posted yesterday,
the attack is exactly the same as before, except that it requires fewer
signatures for full recovery: about 350,000 now vs. 2,500,000 originally
(having a centered y makes convergence faster).

I suspect that the convergence failure you were observing was due to not
updating the random seed between successive calls to the sign() function,
which was presumably causing repeated randomness in generated signatures.

Best regards,

--
M. Tibouchi
<mehdi.t...@normalesup.org>
http://www.normalesup.org/~tibouchi/
************************************************************
NTT Social Informatics Laboratories
Abe Research Laboratory
3-9-11 Midori-cho, Musashino-shi, Tokyo, 180-8585, Japan.
************************************************************


> To view this discussion on the web visit https://groups.google.com/a/list.nist.gov/d/msgid/pqc-forum/ME3PR01MB69195F72F15D5DB77071523DD00BA%40ME3PR01MB6919.ausprd01.prod.outlook.com.

Liu, Dongxi (Data61, Marsfield)

unread,
Aug 2, 2023, 11:25:30 PM8/2/23
to Mehdi Tibouchi, Lorenz Panny, pqc-co...@nist.gov, pqc-...@list.nist.gov

Hi M. Tibouchi,

 

Your attack works. Thanks for checking.

 

Regards,

Dongxi Liu

Liu, Dongxi (Data61, Marsfield)

unread,
Aug 3, 2023, 8:46:42 AM8/3/23
to Mehdi Tibouchi, Lorenz Panny, pqc-co...@nist.gov, pqc-...@list.nist.gov

Hi M. Tibouchi,

 

Based on your observation “not updating the random seed”, the following solution passes leakage attack.

 

In the attacked version, we have s = c1*x1+c2*x2 + (c1+c2)*y,  where y is randomly sampled.  Now we let y = y0 + y1, where  y0 is randomly sampled, but y1 is selected from a small number of cases (4 in the test), which can be sampled as part of the private key or derived from private key.  y1 simulates the effect of “not updating the random seed”.  

 

Here is the updated implementation file:

https://gitlab.com/raykzhao/emle-sig/-/blob/main/Extra/sage-code/leakage-v1/impl.c

 

Attack file is the same.

Thanks for your observation and further checking.

 

Regards,

Dongxi Liu

Mehdi Tibouchi

unread,
Aug 4, 2023, 2:59:38 AM8/4/23
to Liu, Dongxi (Data61, Marsfield), Lorenz Panny, pqc-co...@nist.gov, pqc-...@list.nist.gov
Dear Dr. Liu,

I don't think the proposal here is sufficiently fleshed out to evaluate
(and I suspect a detailed analysis might not be a productive use of
anyone's time), but I can still make a few quick remarks.

The same leakage attacks exactly recovers the following elements:

x1 + y* and x2 + y*

where y* = E[y] is the average of the choices for y1. Now, I strongly
suspect that those values are already sufficient to forge signatures, but
since the provided verification algorithm does not match the signature
generation, it is difficult to test for sure. In any case, however, one
can provide an upper bound on how hard it would be to recover y*, and
hence the key, given the above.

Since x1, x2 has integer coefficients, and y* coefficients in (1/4)Z, one
exactly learns the fractional part of y* from the above, and the
conditional entropy of each coefficient knowing the fractional part is
about 1.5 bits. The conditional entropy knowing x1+y* and x2+y* is again
significantly lower (for example, if one coefficient of either is equal
to 6, one learns for sure that the corresponding coefficient of y* is 2);
I haven't done the exact computation, but it might be around 1 bit per
coefficient or less. Overall, a trivial enumeration attack on the n=64
parameter should therefore take time 2^64.

Moreover, one should be able to carry out a meet-in-the-middle attack on
the average of u to further cut the bit complexity in half.

So I am fairly confident that the proposal remains insecure.

Regards,

--
M. Tibouchi
<mehdi.t...@normalesup.org>
http://www.normalesup.org/~tibouchi/
************************************************************
NTT Social Informatics Laboratories
Abe Research Laboratory
3-9-11 Midori-cho, Musashino-shi, Tokyo, 180-8585, Japan.
************************************************************

> To view this discussion on the web visit https://groups.google.com/a/list.nist.gov/d/msgid/pqc-forum/ME3PR01MB69198D38951970B37AF2E268D008A%40ME3PR01MB6919.ausprd01.prod.outlook.com.

Liu, Dongxi (Data61, Marsfield)

unread,
Feb 17, 2024, 2:26:44 AMFeb 17
to Mehdi Tibouchi, Lorenz Panny, pqc-co...@nist.gov, pqc-...@list.nist.gov

Dear Dr. Tibouchi,

 

Thanks for your time and insight to analyse eMLE-Sig last year.  Your analysis is valid.

 

  1. eMLE-Sig has been thoroughly updated to address all attacks. The updated document, sage implementation, and tests are included in the following folder.  No c implementation is available yet.

 

          https://gitlab.com/raykzhao/emle-sig/-/tree/main/Extra/sage-code/Updates2024Feb

 

  1. Addressing leakage attack

The updated signature still incudes the component s = c1*x1 + c2*x2 + ….  But in the updated scheme, c1[i] and c2[i] are entangled.

c1[i] takes the value either  0 or 1.  If c1[i] =0, then c2[i] must be 1. If c1[i] =1, then c2[i] = -1.

More detailed analysis is in the Section 4.1 of the PDF document in the above folder, with leakage.sage to confirm the analysis there.  

 

  1. Addressing lattice-based attack from the following aspects

The dimension parameter n is increased from 64 to 128 for the first security level;  a new secret z shared between h1 and h2 requires h1 and h2 must be embedded in one lattice to attack, leading to a bigger lattice to reduce than before even for the same n. 

The modulus p[d-1] decreases from 2^26 in the submitted version to 6083 in the new version to make a more dense system.  More details are in the PDF document.

 

  1. Sizes of signatures and public keys

The signature sizes become bigger than the submitted version. 

Level 1:  pk size = 416, sig size = 576 bytes

Level 3:  pk size = 672, sig size = 936 bytes

Level 5:  pk size = 896, sig size = 1280 bytes

 

Welcome any comments and questions!

Liu, Dongxi (Data61, Marsfield)

unread,
Apr 2, 2024, 8:12:31 PMApr 2
to pqc-co...@nist.gov, pqc-...@list.nist.gov

Dear All,

 

The c implementation of updated eMLE-Sig 2.0 is available at: https://gitlab.com/raykzhao/emle-sig

The performance evaluation is included in the updated document here:

https://gitlab.com/raykzhao/emle-sig/-/tree/main/Extra/sage-code/Updates2024Feb

 

Thanks for comments.

 

Regards,

Dongxi Liu

 

From: pqc-...@list.nist.gov <pqc-...@list.nist.gov> on behalf of Liu, Dongxi (Data61, Marsfield) <Dongx...@data61.csiro.au>
Date: Saturday, 17 February 2024 at 6:26
pm
To: Mehdi Tibouchi <mehdi.t...@normalesup.org>
Cc: Lorenz Panny <l.s....@tue.nl>, pqc-co...@nist.gov <pqc-co...@nist.gov>, pqc-...@list.nist.gov <pqc-...@list.nist.gov>
Subject: Re: [pqc-forum] Round 1 (Additional Signatures) OFFICIAL COMMENT: eMLE-Sig 2.0

Dear Dr. Tibouchi,

 

Thanks for your time and insight to analyse eMLE-Sig last year.  Your analysis is valid.

 

1.       eMLE-Sig has been thoroughly updated to address all attacks. The updated document, sage implementation, and tests are included in the following folder.  No c implementation is available yet.

 

          https://gitlab.com/raykzhao/emle-sig/-/tree/main/Extra/sage-code/Updates2024Feb

 

2.       Addressing leakage attack

The updated signature still incudes the component s = c1*x1 + c2*x2 + ….  But in the updated scheme, c1[i] and c2[i] are entangled.

c1[i] takes the value either  0 or 1.  If c1[i] =0, then c2[i] must be 1. If c1[i] =1, then c2[i] = -1.

More detailed analysis is in the Section 4.1 of the PDF document in the above folder, with leakage.sage to confirm the analysis there.  

 

3.       Addressing lattice-based attack from the following aspects

The dimension parameter n is increased from 64 to 128 for the first security level;  a new secret z shared between h1 and h2 requires h1 and h2 must be embedded in one lattice to attack, leading to a bigger lattice to reduce than before even for the same n. 

The modulus p[d-1] decreases from 2^26 in the submitted version to 6083 in the new version to make a more dense system.  More details are in the PDF document.

 

4.       Sizes of signatures and public keys

Reply all
Reply to author
Forward
0 new messages