Argon2

While there is no public cryptanalysis applicable to Argon2d, there are two published attacks on the Argon2i function. The first attack is applicable only to the old version of Argon2i, while the second has been extended to the latest version (1.3).[5]

The first attack shows that it is possible to compute a single-pass Argon2i function using between a quarter and a fifth of the desired space with no time penalty, and compute a multiple-pass Argon2i using only N/e (≈ N/2.72) space with no time penalty.[6] According to the Argon2 authors, this attack vector was fixed in version 1.3.[7]

The second attack shows that Argon2i can be computed by an algorithm which has complexity O(n^7/4 log(n)) for all choices of parameters σ (space cost), τ (time cost), and thread-count such that n=σ∗τ.[8] The Argon2 authors claim that this attack is not efficient if Argon2i is used with three or more passes.[7] However, Joël Alwen and Jeremiah Blocki improved the attack and showed that in order for the attack to fail, Argon2i v1.3 needs more than 10 passes over memory.[5]

To address these concerns, RFC9106 recommends using Argon2id to largely mitigate such attacks. [9]

Function Argon2
   Inputs:
      password (P):       Bytes (0..232-1)    Password (or message) to be hashed
      salt (S):           Bytes (8..232-1)    Salt (16 bytes recommended for password hashing)
      parallelism (p):    Number (1..224-1)   Degree of parallelism (i.e. number of threads)
      tagLength (T):      Number (4..232-1)   Desired number of returned bytes
      memorySizeKB (m):   Number (8p..232-1)  Amount of memory (in kibibytes) to use
      iterations (t):     Number (1..232-1)   Number of iterations to perform
      version (v):        Number (0x13)       The current version is 0x13 (19 decimal)
      key (K):            Bytes (0..232-1)    Optional key (Errata: PDF says 0..32 bytes, RFC says 0..232 bytes)
      associatedData (X): Bytes (0..232-1)    Optional arbitrary extra data
      hashType (y):       Number (0=Argon2d, 1=Argon2i, 2=Argon2id)
   Output:
      tag:                Bytes (tagLength)   The resulting generated bytes, tagLength bytes long

   Generate initial 64-byte block H0.
    All the input parameters are concatenated and input as a source of additional entropy.
    Errata: RFC says H0 is 64-bits; PDF says H0 is 64-bytes.
    Errata: RFC says the Hash is H^, the PDF says it's ℋ (but doesn't document what ℋ is). It's actually Blake2b.
    Variable length items are prepended with their length as 32-bit little-endian integers.
   buffer ← parallelism ∥ tagLength ∥ memorySizeKB ∥ iterations ∥ version ∥ hashType
         ∥ Length(password)       ∥ Password
         ∥ Length(salt)           ∥ salt
         ∥ Length(key)            ∥ key
         ∥ Length(associatedData) ∥ associatedData
   H0 ← Blake2b(buffer, 64) //default hash size of Blake2b is 64-bytes

   Calculate number of 1 KB blocks by rounding down memorySizeKB to the nearest multiple of 4*parallelism kibibytes
   blockCount ← Floor(memorySizeKB, 4*parallelism)

   Allocate two-dimensional array of 1 KiB blocks (parallelism rows x columnCount columns)
   columnCount ← blockCount / parallelism;   //In the RFC, columnCount is referred to as q

   Compute the first and second block (i.e. column zero and one ) of each lane (i.e. row)
   for i ← 0 to parallelism-1 do for each row
      Bi[0] ← Hash(H0 ∥ 0 ∥ i, 1024) //Generate a 1024-byte digest
      Bi[1] ← Hash(H0 ∥ 1 ∥ i, 1024) //Generate a 1024-byte digest

   Compute remaining columns of each lane
   for i ← 0 to parallelism-1 do //for each row
      for j ← 2 to columnCount-1 do //for each subsequent column
         //i' and j' indexes depend if it's Argon2i, Argon2d, or Argon2id (See section 3.4)
         i′, j′ ← GetBlockIndexes(i, j)  //the GetBlockIndexes function is not defined
         Bi[j] = G(Bi[j-1], Bi′[j′]) //the G hash function is not defined

   Further passes when iterations > 1
   for nIteration ← 2 to iterations do
      for i ← 0 to parallelism-1 do for each row
        for j ← 0 to columnCount-1 do //for each subsequent column
           //i' and j' indexes depend if it's Argon2i, Argon2d, or Argon2id (See section 3.4)
           i′, j′ ← GetBlockIndexes(i, j)
           if j == 0 then 
             Bi[0] = Bi[0] xor G(Bi[columnCount-1], Bi′[j′])
           else
             Bi[j] = Bi[j] xor G(Bi[j-1], Bi′[j′])

   Compute final block C as the XOR of the last column of each row
   C ← B0[columnCount-1]
   for i ← 1 to parallelism-1 do
      C ← C xor Bi[columnCount-1]

   Compute output tag
   return Hash(C, tagLength)

Function Hash(message, digestSize)
   Inputs:
      message:         Bytes (0..232-1)     Message to be hashed
      digestSize:      Integer (1..232)     Desired number of bytes to be returned
   Output:
      digest:          Bytes (digestSize)   The resulting generated bytes, digestSize bytes long

   Hash is a variable-length hash function, built using Blake2b, capable of generating
   digests up to 232 bytes.

   If the requested digestSize is 64-bytes or lower, then we use Blake2b directly
   if (digestSize <= 64) then
      return Blake2b(digestSize ∥ message, digestSize) //concatenate 32-bit little endian digestSize with the message bytes

   For desired hashes over 64-bytes (e.g. 1024 bytes for Argon2 blocks),
   we use Blake2b to generate twice the number of needed 64-byte blocks,
   and then only use 32-bytes from each block

   Calculate the number of whole blocks (knowing we're only going to use 32-bytes from each)
   r ← Ceil(digestSize/32)-2;

   Generate r whole blocks.
   Initial block is generated from message
   V1 ← Blake2b(digestSize ∥ message, 64);
   Subsequent blocks are generated from previous blocks
   for i ← 2 to r do
      Vi ← Blake2b(Vi-1, 64)
   Generate the final (possibly partial) block
   partialBytesNeeded ← digestSize – 32*r;
   Vr+1 ← Blake2b(Vr, partialBytesNeeded)

   Concatenate the first 32-bytes of each block Vi
   (except the possibly partial last block, which we take the whole thing)
   Let Ai represent the lower 32-bytes of block Vi
   return A1 ∥ A2 ∥ ... ∥ Ar ∥ Vr+1

As of May 2023, OWASP's Password Storage Cheat Sheet recommends that people "use Argon2id with a minimum configuration of 19 MiB of memory, an iteration count of 2, and 1 degree of parallelism."[10]

OWASP recommends that Argon2id should be preferred over Argon2d and Argon2i because it provides a balanced resistance to both GPU-based attacks and side-channel attacks.[10]

OWASP further notes that the following Argon2id options provide equivalent cryptographic strength and simply trade off memory usage for compute workload:[10]

• Memory: 46 MiB, Iterations: 1, Parallelism: 1
• Memory: 19 MiB, Iterations: 2, Parallelism: 1
• Memory: 12 MiB, Iterations: 3, Parallelism: 1
• Memory: 9 MiB, Iterations: 4, Parallelism: 1
• Memory: 7 MiB, Iterations: 5, Parallelism: 1

• Argon2 source code repository on Github
• Argon2 specification
• Password Hashing Competition
• Uni.Lu Argon2 Page
• Balloon Hashing: A Memory-Hard Function Providing Provable Protection Against Sequential Attacks
• RFC 9106 Argon2 Memory-Hard Function for Password Hashing and Proof-of-Work Applications