PyTorch/CSPRNG

CSPRNG is a PyTorch C++/CUDA extension that provides cryptographically secure pseudorandom number generators for PyTorch.

Design

CSPRNG generates a random 128-bits key on CPU using one of its generators and runs AES128 in CTR mode mode either on CPU or CUDA to generate random 128 bits state and apply transformation function to map it to target tensor values. This approach is based on Parallel Random Numbers: As Easy as 1, 2, 3(John K. Salmon, Mark A. Moraes, Ron O. Dror, and David E. Shaw, D. E. Shaw Research). It makes CSPRNG both crypto-secure and parallel on CUDA and CPU.

Advantages:

• A user can choose either seed-based(for testing) or random device based(fully crypto-secure) generators
• One generator instance for both CPU and CUDA tensors(because the encryption key is always generated on CPU)
• CPU random number generation is also parallel(unlike default PyTorch CPU generator)

Featues

CSPRNG exposes two methods to create crypto-secure and non-crypto-secure PRNGs:

Method to create PRNG	Is crypto-secure?	Has seed?	Underlying implementation
create_random_device_generator(token: string=None)	yes	no	See std::random_device and it's constructor. The implementation in libstdc++ expects token to name the source of random bytes. Possible token values include "default", "rand_s", "rdseed", "rdrand", "rdrnd", "/dev/urandom", "/dev/random", "mt19937", and integer string specifying the seed of the mt19937 engine. (Token values other than "default" are only valid for certain targets.) If token=None then constructs a new std::random_device object with an implementation-defined token.
create_mt19937_generator(seed: int=None)	no	yes	See std::mt19937 and it's constructor. Constructs a mersenne_twister_engine object, and initializes its internal state sequence to pseudo-random values. If seed=None then seeds the engine with default_seed.

The following list of methods supports all forementioned PRNGs:

Kernel	CUDA	CPU
random_()	yes	yes
random_(to)	yes	yes
random_(from, to)	yes	yes
uniform_(from, to)	yes	yes
normal_(mean, std)	yes	yes
cauchy_(median, sigma)	yes	yes
log_normal_(mean, std)	yes	yes
geometric_(p)	yes	yes
exponential_(lambda)	yes	yes

How to build

Since CSPRNG is C++/CUDA extension it uses setuptools, just run python setup.py install to build and install it.

How to use

import torch
import torch_csprng as csprng

# Create crypto-secure PRNG from /dev/urandom:
urandom_gen = csprng.create_random_device_generator('/dev/urandom')

# Create empty boolean tensor on CUDA and initialize it with random values from urandom_gen:
print(torch.empty(10, dtype=torch.bool, device='cuda').random_(generator=urandom_gen))
tensor([ True, False, False,  True, False, False, False,  True, False, False],
       device='cuda:0')

# Create empty int16 tensor on CUDA and initialize it with random values in range [0, 100) from urandom_gen:
print(torch.empty(10, dtype=torch.int16, device='cuda').random_(100, generator=urandom_gen))
tensor([59, 20, 68, 51, 18, 37,  7, 54, 74, 85], device='cuda:0',
       dtype=torch.int16)

# Create non-crypto-secure MT19937 PRNG:
mt19937_gen = csprng.create_mt19937_generator()

print(torch.empty(10, dtype=torch.int64, device='cuda').random_(torch.iinfo(torch.int64).min, to=None, generator=mt19937_gen))
tensor([-7584783661268263470,  2477984957619728163, -3472586837228887516,
        -5174704429717287072,  4125764479102447192, -4763846282056057972,
         -182922600982469112,  -498242863868415842,   728545841957750221,
         7740902737283645074], device='cuda:0')

# Create crypto-secure PRNG from default random device:
default_device_gen = csprng.create_random_device_generator()

print(torch.randn(10, device='cuda', generator=default_device_gen))
tensor([ 1.2885,  0.3240, -1.1813,  0.8629,  0.5714,  2.3720, -0.5627, -0.5551,
        -0.6304,  0.1090], device='cuda:0')

# Create non-crypto-secure MT19937 PRNG with seed
mt19937_gen = csprng.create_mt19937_generator(42)

print(torch.empty(10, device='cuda').geometric_(p=0.2, generator=mt19937_gen))
tensor([ 7.,  1.,  8.,  1., 11.,  3.,  1.,  1.,  5., 10.], device='cuda:0')

print(torch.empty(10, device='cuda').geometric_(p=0.2, generator=mt19937_gen))
tensor([ 1.,  1.,  1.,  6.,  1., 13.,  5.,  1.,  3.,  4.], device='cuda:0')

print(torch.empty(10, device='cuda').geometric_(p=0.2, generator=mt19937_gen))
tensor([14.,  5.,  4.,  5.,  1.,  1.,  8.,  1.,  7., 10.], device='cuda:0')

# Recreate MT19937 PRNG with the same seed
mt19937_gen = csprng.create_mt19937_generator(42)

print(torch.empty(10, device='cuda').geometric_(p=0.2, generator=mt19937_gen))
tensor([ 7.,  1.,  8.,  1., 11.,  3.,  1.,  1.,  5., 10.], device='cuda:0')

print(torch.empty(10, device='cuda').geometric_(p=0.2, generator=mt19937_gen))
tensor([ 1.,  1.,  1.,  6.,  1., 13.,  5.,  1.,  3.,  4.], device='cuda:0')

print(torch.empty(10, device='cuda').geometric_(p=0.2, generator=mt19937_gen))
tensor([14.,  5.,  4.,  5.,  1.,  1.,  8.,  1.,  7., 10.], device='cuda:0')