GitHub - pytorch/csprng: Cryptographically secure pseudorandom number generators for PyTorch (original) (raw)

PyTorch/CSPRNG

torchcsprng is a PyTorch C++/CUDA extension that provides:

Design

torchcsprng generates a random 128-bit key on CPU using one of its generators and runsAES128 in CTR modeeither on CPU or on GPU using CUDA to generate a random 128 bit state and apply a 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 torchcsprng both crypto-secure and parallel on CUDA and CPU.

Advantages:

• The 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 the default PyTorch CPU generator)

Features

torchcsprng 0.2.0 exposes new API for tensor encryption/decryption. Tensor encryption/decryption API is dtype agnostic, so a tensor of any dtype can be encrypted and the result can be stored to a tensor of any dtype. An encryption key also can be a tensor of any dtype. Currently torchcsprng supports AES cipher with 128-bit key in two modes: ECB and CTR.

• torchcsprng.encrypt(input: Tensor, output: Tensor, key: Tensor, cipher: string, mode: string)

• input tensor can be any CPU or CUDA tensor of any dtype and size in bytes(zero-padding is used to make its size in bytes divisible by block size in bytes)
• output tensor can have any dtype and the same device as input tensor and the size in bytes rounded up to the block size in bytes(16 bytes for AES 128)
• key tensor can have any dtype and the same device as input tensor and size in bytes equal to 16 for AES 128
• cipher currently can be only one supported value "aes128"
• mode currently can be either "ecb" or "ctr"

• torchcsprng.decrypt(input: Tensor, output: Tensor, key: Tensor, cipher: string, mode: string)

• input tensor can be any CPU or CUDA tensor of any dtype with size in bytes divisible by the block size in bytes(16 bytes for AES 128)
• output tensor can have any dtype but the same device as input tensor and the same size in bytes as input tensor
• key tensor can have any dtype and the same device as input tensor and size in bytes equal to 16 for AES 128
• cipher currently can be only one supported value "aes128"
• mode currently can be either "ecb" or "ctr"

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

The following list of methods supports all forementioned PRNGs:

• the calculations are done on CPU and the result is copied to CUDA

Installation

CSPRNG works with Python 3.6-3.9 on the following operating systems and can be used with PyTorch tensors on the following devices:

The following is the corresponding CSPRNG versions and supported Python versions.

Binary Installation

Anaconda:

pip:

Nightly builds:

Anaconda:

pip:

From Source

torchcsprng is a Python C++/CUDA extension that depends on PyTorch. In order to build CSPRNG from source it is required to have Python(>=3.7) with PyTorch(>=1.8.0) installed and C++ compiler(gcc/clang for Linux, XCode for macOS, Visual Studio for MS Windows). To build torchcsprng you can run the following:

By default, GPU support is built if CUDA is found and torch.cuda.is_available() is True. Additionally, it is possible to force building GPU support by setting the FORCE_CUDA=1 environment variable, which is useful when building a docker image.

Getting Started

The torchcsprng API is available in torchcsprng module:

import torch import torchcsprng 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:

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:

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() 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() 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) 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')

Recreate MT19937 PRNG with the same seed:

mt19937_gen = csprng.create_mt19937_generator(42) 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')

Contributing

We appreciate all contributions. If you are planning to contribute back bug-fixes, please do so without any further discussion. If you plan to contribute new features, utility functions or extensions, please first open an issue and discuss the feature with us.

License

torchcsprng is BSD 3-clause licensed. See the license file here

Terms of Use

Privacy Policy

Copyright © 2020 Meta Platforms, Inc