Laplace and exponential noise distributions about opacus HOT 5 CLOSED

At least for exponential and normal distribution there appears to be an alternative approach.

One can create a Tensor and fill it with numbers from the given distribution in-place, as described here.
The in-place sampling even allows a generator, so the secure_generator (now called self.random_number_generator) created in the privacy_engine could be utilized and therefore urandom could be used as a source.
The device can be given to a tensor on declaration as well, so this appears to solve both problems.

However this approach lacks Laplace, again.
It is really problematic that the distributions are not standardized so that they behave the same (regarding parameters and configurations) no matter where we call them in the torch libs.

from opacus.

Please let me know if I missed something.

from opacus.

I updated our _generate_noise method to use a tensor for scale and rate to pass on the device.

So the only thing missing is the generator support. Hoping that our proposal in pytorch/pytorch#45115 gets addressed and we will be able to add it later and therefore use urandom as a source.

For our research purposes however this should be sufficient.

def _generate_noise(self, max_norm, parameter):
    if self.noise_multiplier > 0:
        mean = 0
        scale_scalar = self.noise_multiplier * max_norm

        scale = torch.full(size=parameter.grad.shape, fill_value=scale_scalar, dtype=torch.float32, device=self.device)

        if self.noise_type == "gaussian":
            dist = torch.distributions.normal.Normal(mean, scale)
        elif self.noise_type == "laplacian":
            dist = torch.distributions.laplace.Laplace(mean, scale)
        elif self.noise_type == "exponential":
            rate = 1 / scale
            dist = torch.distributions.exponential.Exponential(rate)
        else:
            dist = torch.distributions.normal.Normal(mean, scale)

        noise = dist.sample()

        return noise
    return 0.0

from opacus.

Hi @pepper-jk !

It looks like this is blocked on the PyTorch side, but there are a few comments I'd like to add: in general, while I agree that it would be very simple for us to extend support for more distributions, we are hesitant to do that because we don't see a use case where you would want to use an exponential or Laplace distribution for DP-SGD.

That being said, you should absolutely be able to plug in whatever RNG you want without you having to edit our lib or upstream a PR to us necessarily, just like you don't need to commit a new nn.Module you wrote to PyTorch every time. We are sorry this is not yet very doable, but we are working on making this lib a lot more modular for research for this exact reason :)

For now, I would recommend either editing in-place or monkeypatching in your script:

privacy_engine = PrivacyEngine(...)
privacy_engine._generate_noise = my_generate_noise

from opacus.

Just for the record, We since published a small library we extended to create exponantial and laplace noise using opacus.
In case anyone else needs this.

We are using it in the code for our Label Leakage from Gradients paper.

from opacus.