Performing the attack on a GPU device?

[kfarivar]

Hello
I wanted to know if it is possible to run attacks on a gpu ?

As far as I looked into the code the generate methods for attacks only accept np.ndarray as inputs and there is no parameter when initializing an attack to set a GPU device.
any help is appreciated.

[beat-buesser]

Hi @kfarivar Thank you very much for using ART! Yes, of course you can use GPUs. Setting a specific device depends on the deep learning framework you are using. Which framework are you using?

[beat-buesser]

For PyTorch, for example PyTorchClassifier accepts an argument device_type that can be either gpu or cpu. For gpu it will select the current device.

[awskkzl]

Hi @beat-buesser, is there any way to make the attack towards SklearnClassifier be running on gpu?

[beat-buesser]

Hi @awskkzl I think the answer is no for SklearnClassifier, not because of the attack or ART but because it seems that sklearn is not supporting GPUs in the near future: https://scikit-learn.org/stable/faq.html#will-you-add-gpu-support

[le000043]

Hello @beat-buesser, is GPU the default device on TensorFlowV2Classifier? Thanks

[beat-buesser]

Hi @le000043 With TensorFlow you can set specific devices outside of ART following https://www.tensorflow.org/guide/gpu. Without any specific settings TensorFlow will automatically run on a device if one is available.

[UsamaShakoor]

hello,
i am unable to compute GraphiteBlackBox attack on gpu.
My implementation:
from art.estimators.classification import PyTorchClassifier
from art.attacks.evasion import GRAPHITEBlackbox

classifier = PyTorchClassifier(model=model, device_type="cuda",loss=nn.CrossEntropyLoss(), input_shape= ([10,3,112, 112]), nb_classes=8)

noise_size = (112,112)
net_size = (112,112)
attack = GRAPHITEBlackbox(classifier=classifier, noise_size=noise_size, net_size=net_size,num_xforms_mask=50, num_xforms_boost=50, num_boost_queries=500)

image_adv = attack.generate(x=x_test_class_0, y=y_test_class_0, x_tar=x_tar_replicated)

Where
x=x_test_class_0, y=y_test_class_0, x_tar=x_tar_replicated are all pytorch tensors

which generates following error:

TypeError Traceback (most recent call last)
in <cell line: 18>()
16
17
---> 18 image_adv = attack.generate(x=x_test_class_0, y=y_test_class_0, x_tar=x_tar_replicated)

2 frames
/usr/local/lib/python3.10/dist-packages/art/attacks/evasion/graphite/graphite_blackbox.py in generate(self, x, y, **kwargs)
241 clip_min, clip_max = self.estimator.clip_values
242 else:
--> 243 clip_min, clip_max = np.min(x), np.max(x)
244
245 # target initialization image

/usr/local/lib/python3.10/dist-packages/numpy/core/fromnumeric.py in min(a, axis, out, keepdims, initial, where)
2951 6
2952 """
-> 2953 return _wrapreduction(a, np.minimum, 'min', axis, None, out,
2954 keepdims=keepdims, initial=initial, where=where)
2955

/usr/local/lib/python3.10/dist-packages/numpy/core/fromnumeric.py in _wrapreduction(obj, ufunc, method, axis, dtype, out, **kwargs)
84 return reduction(axis=axis, dtype=dtype, out=out, **passkwargs)
85 else:
---> 86 return reduction(axis=axis, out=out, **passkwargs)
87
88 return ufunc.reduce(obj, axis, dtype, out, **passkwargs)

TypeError: min() received an invalid combination of arguments - got (out=NoneType, axis=NoneType, ), but expected one of:

• ()
• (Tensor other)
• (int dim, bool keepdim)
  didn't match because some of the keywords were incorrect: out, axis
• (name dim, bool keepdim)
  didn't match because some of the keywords were incorrect: out, axis

Any help is appreciated.

[beat-buesser]

Hi @UsamaShakoor To run on GPUs, try setting device_type="gpu" in PyTorchClassifier.

[UsamaShakoor]

Hi, @beat-buesser
Even with device_type="gpu", I'm still encountering the same error. However, if I convert x=x_test_class_0, y=y_test_class_0, and x_tar=x_tar_replicated to numpy arrays, then this error disappears. But in that case, it doesn't run on the GPU.
I want to speedup the process

[beat-buesser]

I took a closer look and it looks like GRAPHITEBlackbox is not implemented with support for GPUs. Are you able to reproduce the original paper?