Sound-Randomized-Smoothing

Code for the paper Sound Randomized Smoothing in Floating-Point Arithmetics https://arxiv.org/abs/2207.07209

It turns out that the standard implementation of randomized smoothing suffers from floating point errors. This sampler fixes the problem.

The provided script contains normal distribution certification procedure for $\ell_2$ certified robustness discussed in the paper.

the usage is as follows:

from certifier import Certifier

Cert = Certifier(sigma=0.5)
certificates = Cert.certify(model, dataset='cifar10')

where model is a base classifier accepting an image tensors in the form NCHW

Additional parameters of Certifier and certify are documented in the script.