When the standard was written in the year 2000 the recommended minimum number of iterations was 1,000, but the parameter is intended to be increased over time as CPU speeds increase. A Kerberos standard in 2005 recommended 4,096 iterations;[1]Apple reportedly used 2,000 for iOS 3, and 10,000 for iOS 4;[4] while LastPass in 2011 used 5,000 iterations for JavaScript clients and 100,000 iterations for server-side hashing.[5] In 2023, OWASP recommended to use 600,000 iterations for PBKDF2-HMAC-SHA256 and 210,000 for PBKDF2-HMAC-SHA512.[6]
Algorithmic representation of the iterative process of the Password-Based Key Derivation Function 2.
Having a salt added to the password reduces the ability to use precomputed hashes (rainbow tables) for attacks, and means that multiple passwords have to be tested individually, not all at once. The public key cryptography standard recommends a salt length of at least 64 bits.[7] The US National Institute of Standards and Technology recommends a salt length of at least 128 bits.[8]
Key derivation process
The PBKDF2 key derivation function has five input parameters:[9]
DK = PBKDF2(PRF, Password, Salt, c, dkLen)
where:
PRF is a pseudorandom function of two parameters with output length hLen (e.g., a keyed HMAC)
Password is the master password from which a derived key is generated
dkLen is the desired bit-length of the derived key
DK is the generated derived key
Each hLen-bit block Ti of derived key DK, is computed as follows (with + marking string concatenation):
DK = T1 + T2 + ⋯ + Tdklen/hlen
Ti = F(Password, Salt, c, i)
The function F is the xor (^) of c iterations of chained PRFs. The first iteration of PRF uses Password as the PRF key and Salt concatenated with i encoded as a big-endian 32-bit integer as the input. (Note that i is a 1-based index.) Subsequent iterations of PRF use Password as the PRF key and the output of the previous PRF computation as the input:
PBKDF1 had a simpler process: the initial U (called T in this version) is created by PRF(Password + Salt), and the following ones are simply PRF(Uprevious). The key is extracted as the first dkLen bits of the final hash, which is why there is a size limit.[9]
HMAC collisions
PBKDF2 has an interesting property when using HMAC as its pseudo-random function. It is possible to trivially construct any number of different password pairs with collisions within each pair.[10] If a supplied password is longer than the block size of the underlying HMAC hash function, the password is first pre-hashed into a digest, and that digest is instead used as the password. For example, the following password is too long:
will generate the same derived key bytes (17EB4014C8C461C300E9B61518B9A18B). These derived key collisions do not represent a security vulnerability – as one still must know the original password in order to generate the hash of the password.[11]
Alternatives to PBKDF2
One weakness of PBKDF2 is that while its number of iterations can be adjusted to make it take an arbitrarily large amount of computing time, it can be implemented with a small circuit and very little RAM, which makes brute-force attacks using application-specific integrated circuits or graphics processing units relatively cheap.[12] The bcrypt password hashing function requires a larger amount of RAM (but still not tunable separately, i.e. fixed for a given amount of CPU time) and is significantly stronger against such attacks,[13] while the more modern scrypt key derivation function can use arbitrarily large amounts of memory and is therefore more resistant to ASIC and GPU attacks.[12]
To limit a brute-force attack, it is possible to make each password attempt require an online interaction, without harming the confidentiality of the password. This can be done using an oblivious pseudorandom function to perform password-hardening.[16] This can be done as alternative to, or as an additional step in, a PBKDF.