2 unstable releases
0.2.0 | Sep 6, 2022 |
---|---|
0.1.0 | May 27, 2022 |
#1419 in Algorithms
46KB
261 lines
fast-erasure-shake-rng
A cryptographically secure fast erasure (i.e. forward secure) pseudo-random number generator, based on the sponge/duplex construction and the keccak-f permutation.
Prioritizes security above performance (speed).
Usage
To create an instance of the RNG, preferably use RngState::new_from_getrandom
. You can hash in additional data using RngState::seed
if wanted. To generate random data use RngState::fill_random_bytes
to fill a buffer with random bytes, or RngState::get_random_bytes
to obtain an array filled with random bytes. The RNG can always be reseeded (if you want backward security) using RngState::seed_with_getrandom
.
Examples
Basic usage:
use fast_erasure_shake_rng::RngState;
let mut rng = RngState::new_from_getrandom().unwrap();
let key = rng.get_random_bytes::<32>();
Reseeding for backward security:
use fast_erasure_shake_rng::RngState;
let mut rng = RngState::new_from_getrandom().unwrap();
let key1 = rng.get_random_bytes::<32>();
// do other things ...
// later:
rng.seed_with_getrandom().unwrap();
let key2 = rng.get_random_bytes::<1000>();
Hashing in additional data, to reduce reliance on the OS RNG:
use fast_erasure_shake_rng::RngState;
let mut rng = RngState::new_from_getrandom().unwrap();
let mut user_dice_rolls = String::new();
std::io::stdin().read_line(&mut user_dice_rolls).unwrap();
rng.seed(user_dice_rolls.as_ref());
let key = rng.get_random_bytes::<32>();
Determinism & Portability
This PRNG is deterministic, meaning that it gives the same output when seeded with the same input(s). Therefore it is necessary to seed it with a non-deterministic source of randomness. The RngState::new_from_getrandom
method crates an instance of the PRNG seeded with randomness obtained from the OS RNG (using the getrandom
crate).
The PRNG is not portable/reproducible though, meaning that the output given the same seeding material may differ between platforms and versions. In particular, the output depends on the targets endianness and the version number of this crate. If you need a PRNG that is deterministic and portable, and don’t need the ability to reseed, use a standard XOF, like SHAKE256.
Crate Features
getrandom
(default): Enable dependency on thegetrandom
crate. This enables convenient and secure ways to seed the RNG, e.g.RngState::new_from_getrandom
.rand-core
: Enable dependency on therand_core
crate. This enables implementations of the Rng traits fromrand_core
forRngState
.
RNGs and Cryptography Notes
Attacker controlled entropy sources
There is a widespread idea that adding as many entropy sources as possible to an RNG is always a good idea. Unfortunately, life is not that easy. If an entropy source is attacker controlled, it can actually be used to weaken or break the RNG. This does require the malicious entropy source to know or guess the input from other (earlier added) entropy sources, or know the RNG state (i.e. the randomness pool). While this is not a mild assumption, it is more conceivable for the malicious entropy source to have access to these inputs than for the attacker himself to have them. Therefore it is still a good idea to be a bit careful with adding entropy sources. See e.g. https://blog.cr.yp.to/20140205-entropy.html for more information.
Backward security
This RNG allows for reseeding, i.e. hash extra entropy into the state (after initial use). This would provide backward security: even if an attacker has observed the RNG state in the past, he won’t be able to predict the output of the RNG after the reseed operation has happend. In practice though, if an attacker is able to see the RNG state you have much worse problems then RNG security if you want to perform crypto. Hence there is little to no reason to actually reseed the RNG after some time.
Design
I initially thought that just a sponge/duplex construction over some (secure) permutation, would yield a secure RNG, but unfortunately that construction is not forward secure. The problem is that the used permutation, keccak-f[1600] is efficiently invertible. Hence if no reseed is performed after obtaining random bytes, an attacker that obtained the final state of the RNG could just apply the inverse permutation repeatedly to find all the bytes that were emitted since the last reseed.
What makes the sponge construction still secure for a hash function or XOF is that the bytes in the capacity part of the state are never published, unlike what we have to assume for forward security of the RNG. What would solve the problem is to zeroize the capacity part of the state after every permutation application. But then the RNG doesn’t maintain a secret entropy pool: given only one rate sized output, one could compute all future values until a reseed. Hence we actually need two separate capacity parts in the state: one which is always (that is, every time after outputting random bytes) zeroized for forward secrecy and one that acts as the usual (entropy collecting) state part.
Detailed description
The keccak state consists, as usual, of 1600 bits, divided into 25 lanes of 64 bits each. We divide it into three “areas”:
- An area we call the “rate area”, consisting of the top 9 lanes, therefore sized 576 bits.
- An area we call the “zeroized capacity area”, consisting of the next 8 lanes, therefore sized 512 bits.
- An area we call the “capacity area”, consisting of the last 8 lanes, therefore sized 512 bits.
We now define three basic actions on this state, which will serve as the building blocks for all other (user facing) operations:
- A basic action we call “input”. First xor 576 bits of input data into the “rate area” of the state. Then apply keccak-f to the state.
- A basic action we call “initial-output”. First output the bytes in the “rate area” as random output bytes. Then apply keccak-f.
- A basic action we call “intermediate-output”. First output the bytes in the “rate area” and the “zeroized capacity area” as random output bytes. Then apply keccak-f to the state.
- A basic action we call “make-forward-secure”. Zeroize (i.e., fill with zero/null bytes) the “zeroized capacity area”.
The first basic actions is used to absorb entropy from inputs into the state. The next two are used squeeze output from the state. The action “make-forward-secure” creates a point of forward-security: if the state is leaked after this action then an attacker won’t be able to infer inputs to or outputs from the RNG performed before this action.
In a diagram:
Basic action 1: input State: Basic action 2: initial-output
┌────────┐
│Rate │
Xor input │ │ Output
───────────►│9 lanes ├───────────►
│ │
│ │
│ │
│ │
│ │
│ │
├────────┤
│Zeroized│
Leave alone │capacity│ Leave alone
│ │
│8 lanes │
│ │
│ │
│ │
│ │
├────────┤
│Capacity│
Leave alone │ │ Leave alone
│8 lanes │
│ │
│ │
│ │
Then apply Keccak-f1600 to "state". │ │ Then apply Keccak-f1600 to "state".
│ │
└────────┘
Basic action 3: intermediate-output State: Basic action 4: make-forward-secure
┌────────┐
│Rate │
Output │ │ Leave alone
◄───────────┤9 lanes │
│ │
│ │
│ │
│ │
│ │
│ │
├────────┤
│Zeroized│
Output │capacity│ Zeroize
◄───────────┤ │◄───────────
│8 lanes │
│ │
│ │
│ │
│ │
├────────┤
│Capacity│
Leave alone │ │ Leave alone
│8 lanes │
│ │
│ │
│ │
Then apply Keccak-f1600 to "state". │ │
│ │
└────────┘
Inputting data into the the RNG works as follows: The data is padded to a multiple of 576 bits (72 bytes) using simple 10*1 bit padding. For each 576 bit chunk of input, the basic action “input” is executed with this chunk as input. Finally, the basic action “make-forward-secure” is performed.
Outputting data from the RNG works as follows: First, the basic action “initial-output” is executed, and the output of it is used as the first part of the output of the RNG. If more data was requested (i.e., more than 72 bytes), then the basic action “intermediate-output” is executed repeatedly, until enough output has been generated. Finally, the basic action “make-forward-secure” is performed.
TODOs
- improve throughput for large requests by utilizing intermediate results in the "zeroized state part"
- add a fast-erasure buffered mode for more efficient frequent small requests
Changelog
See CHANGELOG.md
.
Documentation
The API documentation of fast-erasure-shake-rng
is available at https://docs.rs/secmem-alloc/*/fast_erasure_shake_rng/.
Dependencies
~105–285KB