This is the Very Complicated Credential Service.

The VCCS is en-route to becoming a complete authentication system with at
least a SAML IdP front end, a distributed user database, separate password
hashing servers etc.

Design started from performance requirements, so a nearly complete backend
service for password and OTP validation was the first thing to be
implemented.


Authentication
==============

Pseudo-code describing the authentication process ('|' denotes concatenation) :

  On frontend (SAML IdP, RADIUS server etc, see python-vccs_client) :
  -------------------------------------------------------------------

  credential_id, salt, parameters = load_from_userdb(user_id)
  T1 = (credential_id | plaintext_password)

  // Get rid of plaintext as soon as possible, to avoid leaking it (logs etc.)
  H1 = BCRYPT_PBKDF(T1, salt, parameters)
  send_to_backend(H1, user_id, credential_id)


  On backend (dedicated hashing servers, this is what is implemented here) :
  --------------------------------------------------------------------------

  credential_stored_hash, iterations, salt, key_handle = \
  	load_from_private_database(credential_id)

  T1 = 'A' | user_id | credential_id | H1
  T2 = PBKDF2-HMAC-SHA512(T1, iterations=many, salt)
  local_salt = YHSM_HMAC_SHA1(T2)
  H2 = PBKDF2-HMAC-SHA512(T2, iterations=1, local_salt)

  audit_log(frontend_id, credential_id, trunc(H2), trunc(credential_stored_hash))

  return (H2 == credential_stored_hash)


Rationale of the hashing scheme
===============================

Two algorithms was chosen

  a) to not place all trust in a single algorithm
  b) to increase the customization required by an attacker to attack this
     particular systems. This is probably only of real relevance as counter-
     measure to ASIC based attacks (and maybe FPGA)
  c) to be able to do a part of the hashing as soon as possible (in the
     front ends), while being able to scale hashing computation power easier
     in dedicated backends

bcrypt was chosen since it is generally ranked very high in strength against
brute force attacks (bcrypt does not appear to execute faster on GPUs than
CPUs today, although that will likely change with future GPUs).

PBKDF2-HMAC-SHA512 was chosen because it is very well studied by now, as well
as perhaps being a requirement for some. SHA512 was preferred over other SHA's
because it currently favours the defender since it is pessimising anyone using
only 32 bit operations (such as contemporary GPUs).


Annotations for the hashing scheme
==================================

Authentication client code (see separate project python-vccs_client) :

  T1 = credential_id | plaintext_password

credential_id should be a unique identifier for this particular credential.
An actual user might have many credentials associated with him/her. The
credential_id should never be reused, so a changed password should generate a
new unique credential_id and revoke the old one. This is in order to prevent
an attacker controlling the user database from restoring old (cracked)
credential information. The revocation status is stored in the trusted
credential database only accessible to the backends.

  // Get rid of plaintext as soon as possible, to avoid leaking it (logs etc.)
  H1 = BCRYPT_PBKDF(T1, userdb_stored_salt)

plaintext_password is the password entered in the IdP HTML login form, for
example.

userdb_stored_salt is the bcrypt salt in the current version. It is
read from the user database and should not be sent to the backends.
This way, an intruder on the backends will not get the full bcrypt KDF hashes
to mount a relatively unexpensive offline attack on. The intruder will lack
information about the 128 bits (minimum) salt so cannot feasibly attack the H1
hashes.

The computation of H1 should be performed as soon as possible in the frontend
code, to immediately get rid of the plaintext password - hopefully before
it ends up in a debug log, swap file, core dump or similar.

This pre-hashing step has evolved from standard bcrypt(), to bcrypt() with
the salt removed from what is sent to the backends, to the new construct
bcrypt_pbkdf().

  (http://www.openbsd.org/cgi-bin/cvsweb/src/lib/libutil/bcrypt_pbkdf.c)

The reason to remove the salt is to not give an attacker at the authentication
backends the ability to record H1's for later offline attack. The reason for
moving to the new (and therefor less proven) bcrypt_pbkdf() is to avoid the
standard bcrypt() limitation of only 72 characters significant input -
something highly unintuitively for a complicated credential system.

bcrypt_pbkdf still works with standard cryptographic constructs although code
immaturity and lack of independent implementations is a concern.


Authentication backend code :

  T1 = 'A' | user_id | credential_id | H1

The 'key usage' is 'A' (A for authentication). It is included in T just in
case someone builds, say, a signing service with the same scheme, where users
use the same password. Really, just as a precaution that will perhaps allow
solving some future problem.

user_id is a unique identifier for this user. It might be a username, an
e-mail address, a UUID or some other type of persistent identifier of a specific
user. It is included in the computations on the backends to prevent an
attacker that can modify the user database to re-associate a (known) credential
with another user to gain access to that other users account. This might or
might not be effective, based on properties of the frontends and userdb. If an
attacker can write to the userdb, he can probably reassociate a credential as
well as this user_id to another user, in order to get (at least short term)
access to that other user account.

credential_id is the unique identifier of this particular credential. In the
current implementation, this is simply an integer.

H1 is the bcrypt() hash, but does not include the bcrypt salt (the bcrypt
operation is performed on the front end). It is imperative that H1 is a part
of the local parametrization using the YubiHSM. This prevents an attacker
that gain access to the authentication backend from just using that access
to calculate all the PBKDF2 salts for all the credentials, to later be used
in an offline attack. This way, an attacker can just get the PBKDF2 salts
for users that log in during the time period they have access to the authen-
tication backends.


  T2 = PBKDF2-HMAC-SHA512(T1, iter=many, salt)

This is the really time consuming part of validation. The number of interations
is stored per credential in the private credential database. This allows for
more iterations where deemed appropriate, such as with system administrator
accounts.


  local_salt = YHSM_HMAC_SHA1(T)

local_salt is the result of a keyed HMAC-SHA-1 calculation, done inside the
YubiHSM. It is not computed until after the 'T2 = PBKDF2-HMAC-SHA512' above,
so that an intruder on the authentication backends will have to do as much
work up front to calculate local_salts of credentials. If local_salt was
derived earlier in the process, an intruder could potentially use the
(temporary) access to calculate lots of local_salts for use later in an
offline attack on the full H2 accounts (by sending candidate passwords to
the frontends to get T1 and then computing the local_salt without having
to do the time consuming PBKDF2 in real time).

The key is not possible to extract by an attacker that gains access
to the authentication backend, and the result is 160 bits of salt per
credential, with very good random distribution. If an attacker acquires any
number of H2 hashes (those stored in the credential database), he will have
to attempt on average 2^159 different possible salts per attempted plaintext,
thus in effect multiplying the work factor for the attacker with about 10^47,
which is a really really large number.


  H2 = PBKDF2-HMAC-SHA512(T2, iterations=1, local_salt)

H2 is the final hash, that is compared to what is stored in the credential
database. This second PBKDF2 step is performed to mix in the local_salt
into T2 in a good way - just like scrypt uses PBKDF2 as the final step.


Key management
==============

The HMAC keys used to derive the locally parameterized salt (in the dedicated
hashing servers) from the YubiHSMs should not be used indefinitely.

If local password retention policy mandates users to change password every
three years, and key management policy limits key usage to credentials created
in a six months period, a total of seven six-months keys could be allowed to
co-exist simultaneously in the YubiHSM.

By having seven keys, the user can be authenticated for six months after the
password expires to allow for a password change. After 3.5 years, the password
would have to be reset using some external mechanism.

NIST SP 800-57 has recommendations regarding the lifetime of an "Symmetric
authentication key", saying that it should not be used on Originator systems
(operation: password set) for more than two years, and not on Recipient
systems (operation: password validation) for more than two+three years.

---

Fredrik Thulin <[email protected]>, 2013-06-25