Re: Review of draft-ecc-09
2012-02-21 01:45:51
Hello Marko, thank you for taking time to write this review and for the
effort to implement it. Please check my comments inline:
On 02/18/2012 01:51 PM, Marko Kreen wrote:
Hello,
I tried to implement ECC for pgcrypto, which is crypto module
for PostgreSQL database. And I managed to get it to work,
mainly because of EC module in OpenSSL, which allowed me to be
ignorant of all low-level math details.
Also, I only implemented ECDH, pgcrypto does not do signing.
It uses PGP as a fancy encrypt/decrypt storage format only.
So this is a review of draft-08 by app developer who is ignorant
of EC math and has not read any detailed math/crypto papers...
[ I updated the review with diff from -09. Thanks for taking
my comments on the ref section into account. ]
Sorry, I was busy in recent days and this is the first time I take close
look at your feedback. The credit for -09 should go to Sean Turner, IETF
Security Area director.
5. Supported public key algorithms
Supported public key algorithms are Elliptic Curve Digital
Signature Algorithm (ECDSA), defined in [FIPS 186-2], and Elliptic
Curve Diffie-Hellman (ECDH), defined in section 8.
Note for later: this basically states that section 8 plans to
fully describe ECDH used in OpenPGP.
6. Conversion primitives
The method to convert an EC point to the octet string is defined in
[SEC1]. This specification only defines uncompressed point
format. For convenience, the synopsis of the encoding method is
given below, however, the [SEC1] is the normative source of the
definition.
The point is encoded in MPI format. The content of the MPI is the
following:
B = B0 || x || y
where x and y are coordinates of the point P = (x, y), each encoded
in big endian format and zero-padded to the underlying field size.
*Then*, they are also padded to byte boundary. As this is not mentioned
anywhere, it caused me some confusion, because I assumed they already
are on byte boundary, perhaps even power-of-two.
As it only matters to P-521 keys, the bad assumtions work fine on
P-256 and P-384 keys. (Basically I assumed P-521 uses 512-bit values...)
Note, that what you are suggesting would not work in practice, if I
understand your initial confusion correctly. Note that both x and y are
taken mod a 521 bit prime. They can be exactly 521 bits, but in 50% of
cases, 520 bits. Let's say that either x or y is 521 and another one is
520 bits. How would we know which one of x and y is the longest one?
The algorithm is the following: given the MPI with the point,
* get its size in full bytes ( len=(MPI_bits+7)/8 )
* len = len-1 ( to remove the B0 byte )
* len = len/2 -- the 'len' is the size of x and y in bytes.
B0 is a byte with following values:
value description
0 Point O. In this case there is no x or y octets present.
4 Uncompressed point. x and y of EC point values follow.
Note that point O shall not appear in a public or a private
key. Therefore, the size of the MPI payload is always curve_size*2
+ 3 bits. For example, for "Curve P-256" the point is represented
as a bit string of length 515 bits.
Here it is clearly wrong - it is not (curve_size * 2 + 3), but
(curve_size_padded_to_8 * 2 + 3).
I accept your correction. The "curve_size*2+3" above is only true for
sizes that are even to 8 bits.
Missing detail: In addition to size check, what other validation must
be done when parsing a point? This applies to when reading a public key,
but especially when reading incoming ECDH/ECDSA message. I suggest
adding: "You must check that the point is on curve." here.
I will take it as a suggestion and think how to integrate it, given that
this is a general check for ECC and is not specific to OpenPGP. This
valid concern may be satisfied through sources in References section.
8. EC DH Algorithm (ECDH)
The key wrapping method is based on [RFC3394]. KDF produces the
AES key that is used as KEK as specified in [RFC3394]. Refer to
section 13 for the details regarding the choice of the KEK
algorithm, which MUST be one of three AES algorithms.
This is only place which says that KEK *MUST* be AES-only. (Ignoring 12.2)
This is in conflict with section 12.1 and 13, where non-AES is not
disallowed. I'm not disagreeing with it, just I think it would be good
to clarify why.
It might be good idea to disallow cipher_ids< AES128, but why disallow
Camellia and Twofish? I could imagine that because the algorithm
is not per-message but comes from key is the reason - you may not
know at key generation time all the people who want to send you messages
and what features their software has. But I think it would be
good to put the reason in the doc.
In any case, 12.1 and 13 should be synced with the requirement.
I also see the need for clarity here.
In reference to compatibility, if the code understands ECC, the document
assumes that it must supports AES. This is something OpenPGP RFC 4880
cannot generally rely upon. One of main criteria of this document is
simplicity through limited choices, as I recall, something that Ian G
advocated. It's much easier to write "use AES" than to define "strong
enough" cipher. Suite B and other government standards require AES, so
most application will implement it to comply, leaving us with the
question why bother at all about the non-AES case?
I am leaning toward fixing KEK at AES. There is no backward
compatibility issue here. So, the action item for me is to make this
choice clear.
For convenience, the synopsis of the encoding method is given
below, however, this section, [NIST SP800-56A], and [RFC3394] are
the normative sources of the definition.
Obtain authenticated recipient public key R
Generate ephemeral key pair {v, V=vG}
Compute shared point S = vR;
m = symm_alg_ID || session key || checksum || pkcs5_padding;
curve_OID_len = (byte)len(curve_OID);
Param = curve_OID_len || curve_OID || public_key_alg_ID || 03 ||
01 || KDF_hash_ID || AES_alg_ID for AESKeyWrap ||
"Anonymous Sender " || recipient_fingerprint;
Z_len = key size for AES_alg_ID to be used with AESKeyWrap
Compute Z = KDF( S, Z_len, Param );
Compute C = AESKeyWrap( Z, m ) as per [RFC3394]
VB = convert point V to octet string
Output (MPI(VB) || len(C) || C).
The decryption is the inverse of the method given. Note that the
recipient obtains the shared secret by calculating
S = rV = rvG, where (r,R) is the recipient's key pair.
Consistent with section 5.13 Sym. Encrypted Integrity Protected
Data Packet (Tag 18) of [RFC4880], the MDC SHOULD be used anytime
symmetric key is protected by ECDH.
Missing detail: How to generate v? What requirements it has?
I suggest expanding the second step with:
Generate ephemeral key pair {v, V} where V=vG and v is random number
in range 0< v< n (n - curve modulus)
This detail is also related to a question of external reference, but I
am glad you raised this question. FIPS 186-3 defines two methods in B.4
on p.61. The idea is this:
B.4.1: if you do "v = v' mod order", the input v' must be 64 bits longer
than "order".
B.4.2: or you can repeatedly getting random bits until "v' < order"
(probably too wasteful for the entropy)
As I noted earlier 8) claims to define ECDH as used in OpenPGP,
and it does - I managed to implement ECDH without digging in those
"real" specs. That was the only point that I needed to look up.
I think is worth adding here, mostly because it simple.
So far there are two general ECC methods that you found helpful. I will
think what's the best way to integrate them.
9. Encoding of public and private keys
As an implementation note, observe that the ECDH public key fields
are the super-set of the ECDH key fields.
First one should be ECDSA.
Thank you, will correct this.
13. Security Considerations
SHA-1 MUST NOT be used for ECDSA or with KDF in ECDH method.
But MD5 can? How about RIPEMD160? Why single out SHA1?
Because it's a hardwired default choice of RFC 4880 for the hash
algorithm and so it's implicitly assumed to be present in the hash
preference list. I agree that "SHA-1 or weaker" simply looks better.
Final note: the section 8 was quite successful at describing ECDH,
how hard would it be to have same level of description of ECDSA here?
At least I would like to see packet format here, even if the
algorithm is not described.
Marko, did you realize that it happens that the signature format for
ECDSA is identical to DSA? The (r,s) signature pair consists of (true)
MPIs in both cases. The ECDSA key is defined in section 9. Overall,
ECDSA for OpenPGP is much simpler because it has no complexities like
key wrapping. With this in mind, what would you like to see added to
properly implement ECDSA.
The simlilarity ECDSA with DSA in in OpenPGP is a good thing because
signature keys are the foundation for interoperability. For example,
it's the top signature key that validates the key preferences and holds
subkeys.
Thank you, again, for your comments.
|
|