Storing passwords for user authentication purposes in plaintext is the simplest but least secure method; when an attacker compromises the database in which the passwords are stored, they immediately gain access to all of them. Often passwords are reused among multiple services or machines, meaning once a password to a single service is known an attacker has a substantial head start on attacking other machines.
The general approach is to store, instead of the password, the output of a one way function of the password. Upon receiving an authentication request, the authenticating party can recompute the one way function and compare the value just computed with the one that was stored. If they match, then the authentication request succeeds. But when an attacker gains access to the database, they only have the output of the one way function, not the original password.
Common hash functions such as SHA-256 are one way, but used alone they have problems for this purpose. What an attacker can do, upon gaining access to such a stored password database, is hash common dictionary words and other possible passwords, storing them in a list. Then he can search through his list; if a stored hash and an entry in his list match, then he has found the password. Even worse, this can happen offline: an attacker can begin hashing common passwords days, months, or years before ever gaining access to the database. In addition, if two users choose the same password, the one way function output will be the same for both of them, which will be visible upon inspection of the database.
There are two solutions to these problems: salting and iteration. Salting refers to including, along with the password, a randomly chosen value which perturbs the one way function. Salting can reduce the effectiveness of offline dictionary generation, because for each potential password, an attacker would have to compute the one way function output for all possible salts. It also prevents the same password from producing the same output, as long as the salts do not collide. Choosing n-bit salts randomly, salt collisions become likely only after about 2:sup:(n/2) salts have been generated. Choosing a large salt (say 80 to 128 bits) ensures this is very unlikely. Note that in password hashing salt collisions are unfortunate, but not fatal - it simply allows the attacker to attack those two passwords in parallel easier than they would otherwise be able to.
The other approach, iteration, refers to the general technique of forcing multiple one way function evaluations when computing the output, to slow down the operation. For instance if hashing a single password requires running SHA-256 100,000 times instead of just once, that will slow down user authentication by a factor of 100,000, but user authentication happens quite rarely, and usually there are more expensive operations that need to occur anyway (network and database I/O, etc). On the other hand, an attacker who is attempting to break a database full of stolen password hashes will be seriously inconvenienced by a factor of 100,000 slowdown; they will be able to only test at a rate of .0001% of what they would without iterations (or, equivalently, will require 100,000 times as many zombie botnet hosts).
Memory usage while checking a password is also a consideration; if the computation requires using a certain minimum amount of memory, then an attacker can become memory-bound, which may in particular make customized cracking hardware more expensive. Some password hashing designs, such as scrypt, explicitly attempt to provide this. The bcrypt approach requires over 4 KiB of RAM (for the Blowfish key schedule) and may also make some hardware attacks more expensive.
Botan provides two techniques for password hashing, bcrypt and passhash9.
Bcrypt is a
password hashing scheme originally designed for use in OpenBSD, but numerous
other implementations exist. It is made available by including
It has the advantage that it requires a small amount (4K) of fast RAM to compute, which can make hardware password cracking somewhat more expensive.
Bcrypt provides outputs that look like this:
Due to the design of bcrypt, the password is effectively truncated at 72
characters; further characters are ignored and do not change the hash. To
support longer passwords, one common approach is to pre-hash the password
with SHA-256, then run bcrypt using the hex or base64 encoding of the hash as
the password. (Many bcrypt implementations truncate the password at the first
NULL character, so hashing the raw binary SHA-256 may cause problems. Botan’s
bcrypt implementation will hash whatever values are given in the
std::string including any embedded NULLs so this is not an issue, but
might cause interop problems if another library needs to validate the
generate_bcrypt(const std::string &password, RandomNumberGenerator &rng, uint16_t work_factor = 12, char bcrypt_version = "a")¶
Takes the password to hash, a rng, and a work factor. The resulting password hash is returned as a string.
Higher work factors increase the amount of time the algorithm runs, increasing the cost of cracking attempts. The increase is exponential, so a work factor of 12 takes roughly twice as long as work factor 11. The default work factor was set to 10 up until the 2.8.0 release.
It is recommended to set the work factor as high as your system can tolerate (from a performance and latency perspective) since higher work factors greatly improve the security against GPU-based attacks. For example, for protecting high value administrator passwords, consider using work factor 15 or 16; at these work factors each bcrypt computation takes several seconds. Since admin logins will be relatively uncommon, it might be acceptable for each login attempt to take some time. As of 2018, a good password cracking rig (with 8 NVIDIA 1080 cards) can attempt about 1 billion bcrypt computations per month for work factor 13. For work factor 12, it can do twice as many. For work factor 15, it can do only one quarter as many attempts.
Due to bugs affecting various implementations of bcrypt, several different variants of the algorithm are defined. As of 2.7.0 Botan supports generating (or checking) the 2a, 2b, and 2y variants. Since Botan has never been affected by any of the bugs which necessitated these version upgrades, all three versions are identical beyond the version identifier. Which variant to use is controlled by the
The bcrypt work factor must be at least 4 (though at this work factor bcrypt is not very secure). The bcrypt format allows up to 31, but Botan currently rejects all work factors greater than 18 since even that work factor requires roughly 15 seconds of computation on a fast machine.
check_bcrypt(const std::string &password, const std::string &hash)¶
Takes a password and a bcrypt output and returns true if the password is the same as the one that was used to generate the bcrypt hash.
Botan also provides a password hashing technique called passhash9, in
passhash9.h, which is based on PBKDF2.
Passhash9 hashes look like:
This function should be secure with the proper parameters, and will remain in the library for the foreseeable future, but it is specific to Botan rather than being a widely used password hash. Prefer bcrypt.
This password format string (“$9$”) conflicts with the format used for scrypt password hashes on Cisco systems.
generate_passhash9(const std::string &password, RandomNumberGenerator &rng, uint16_t work_factor = 15, uint8_t alg_id = 4)¶
Functions much like
generate_bcrypt. The last parameter,
alg_id, specifies which PRF to use. Currently defined values are 0: HMAC(SHA-1), 1: HMAC(SHA-256), 2: CMAC(Blowfish), 3: HMAC(SHA-384), 4: HMAC(SHA-512)
The work factor must be greater than zero and less than 512. This performs 10000 *
work_factorPBKDF2 iterations, using 96 bits of salt taken from
rng. Using work factor of 10 or more is recommended.
check_passhash9(const std::string &password, const std::string &hash)¶
Functions much like