October 14, 2016

# Memory Hard Functions

Memory-Hard Functions (MHFs) are one of the hot topics in cryptography right now. If you’re not familiar with the term, the gist is as follows. Given time $T$ and space $S$, an optimal MHF is one where $S \times T \in \Omega(n^2)$, where $n$ is the security parameter [1]. This means that it is computationally expensive to compute the function in terms of both time and space. The goal of constructing MHFs is to increase the constant term for $n^2$ to as large as possible. Intuitively, this makes the function expensive to compute in any computer.

Let me elaborate on this point in more detail. Today, the most common use case for MHFs are as password hashing algorithms. To deter dictionary attacks, passwords are (or should be) computed with an additional salt value. This expands the search space for a single password exponentially by the length of the salt. However, this is really only relevant to online dictionary attacks against these functions. Offline dictionary attacks, on the other hand, can be carried out with tremendously more resources and dedicated hardware. Depending on the type of password hashing algorithm used, these attacks can range from trivially cheap to incredibly expensive. Standard cryptographic hash functions are insufficient here since they can be effortlessly computed using special-purpose dedicated hardware. To given an example mentioned in [1], it takes approximately $100,000\times$ more energy to compute a single SHA-256 invocation on a general purpose CPU than on special purpose hardware (e.g., an ASIC) [2]. While it is true that requiring more iterations of the hash function increases the cost of its computation, this is of negligible effect on dedicated platforms.

This is where MHFs come in. If increasing the time dimension does not do the trick, we must look to the space dimension. Since MHFs require a large amount of password independent state to compute, dedicated platforms are much more costly to implement. The attacker must choose to use its resources for more “cores” (or hash execution engines) or more memory. And with a finite amount of resources available, any choice will lead to significantly less computational power to hash passwords. Basically, the large memory requirement ensures that special-purpose hardware can only contain a small number of cores. The attacker must choose to dedicate his or her resources to more space or more computation time, not both. That’s intuitively where the “hardness” comes from.

There are many MHFs floating around the cryptosphere right now. These include scrypt [3], Argon2 [4], Catena [5], and Balloon [1] (to name a few). This is in large part due to the recently concluded Password Hashing Competition [6], which named Argon2 as the winner. I’ll admit I didn’t follow the developments closely over the past couple of years, but I’m trying to make due for lost time now.

# MHF Constructions

Standard MHFs take four inputs and produce a single out. The inputs include the password to hash, a salt, and space and time parameters. The password and salt are what you might expect: strings. The tuning parameters are more interesting. The space parameter $s$ determines how much memory is to be consumed by the MHF. As [1] puts it, the MHF should be easy to compute with $s$ blocks of memory but difficult to do so with anything less. The time parameter $r$ denotes the number of rounds the function iterates before producing its output. Typically, this is tuned to account for what space is available (or not) when computing the function. On systems without a great deal of memory, $r$ is increased to maintain the memory hardness properties.

Most MHFs generally consist of three stages: expand, mix, and output. Their role is clear: expand takes the inputs and blows it up into a large amount of memory, mix iterates over this memory and diffuses the inputs throughout the entire state space, and output produces the final digest. The mix step is where the meat of the algorithm works. And it’s typically designed such that it cannot be parallelized. On multi-core machines, there would be ample room to then execute multiple invocations in simultaneously. This sort of defeats the purpose of the function’s inherent computational hardness. (Especially on systems with many cores, such as GPUs.) Therefore, to remedy this problem, many MHFs specify “parallel” versions. If $F(p, s)$ is the MHF with password $p$ and salt $s$, and $M$ cores are available, then the parallel version $F_M(p, s)$ is:

Basically, $M$ different invocations are combined to produce the final result. This is exactly what the Balloon algorithm does, at least. It’s important to note that the salt is, well, salted for each of the $M$ invocations of $F$. Modifying the password would be silly for a number of reasons. First, at least in the case of Balloon, the salt determines the memory access pattern. If this is unchanged across cores, then it might be possible to precompute this pattern and therefore speed up parallel implementations. By modifying the salt this is effectively stopped. Second, we shouldn’t be messing with the password – the most sensitive input – anyway.

Now let’s move on to some code. Here’s the pseudocode description of Balloon. I’m focusing on it so much because I think it’s simply elegant.

It’s amazingly simple. You can see that the amount of memory depends solely on $s$. The expansion phase is intuitive: feed the password and salt into a cryptographic hash function (necessary for one-wayness) to initialize the first block of memory. Then, initialize the remaining blocks based on a monotonically increasing counter and the value of the previous block. The mixing step is also very clear. There are $st$ iterations (which points to our earlier security definition where $s \times t \in \Omega(n^2)$), and during each iteration the following occurs. Every block $B$ is advanced based on the previous (iteration) block and the current value of $B$ (lines 18-19). Then, a random selection of other blocks are mixed into $B$. The selection of these other blocks depends on the iteration count and salt. (This is why memory accesses are dependent on the salt.) When all is said and done, the last block is chosen as the output.

It’s so simple in fact that I was able to whip it together in Go in about 30 minutes. Here’s the program (the full repository is here. You may notice that I don’t implement the pseudorandom index generator faithfully. The Balloon code does a better job of this [7].

Does this behave as expected? I profile the runtime and memory usage of the program for various values of $s$ and $r$. These values are plotted below as a function of both $s$ and $r$. As one might expect, they both increase as these variables increase. But the space parameter certainly has more control over the runtime than the time parameter.