IndexError on pre-fitted SVC on FastGradientMethod

[melikedem]

Hello,
I have a very simple question.
I have a basic setup with a trained SVC and I would like to try FastGradientMethod on it. I create SklearnClassifier with the following line based on my sklearn SVC:
classifier = SklearnClassifier(model=svc)

However I get "IndexError: arrays used as indices must be of integer (or boolean) type" when I try calling FastGradientMethod.
I do not want to re-fit my SklearnClassifier as my SVC is already fitted.

I would appreciate any help.
Thank you

code snippet: (I can share more if needed)
from art.attacks.evasion import FastGradientMethod
from art.estimators.classification import SklearnClassifier

classifier = SklearnClassifier(model=svc)
attack = FastGradientMethod(estimator=classifier, eps=0.2)
x_test_adv = attack.generate(x=scaledXTest)

Error:
IndexError Traceback (most recent call last)
in
4 classifier = SklearnClassifier(model=svc)
5 attack = FastGradientMethod(estimator=classifier, eps=0.2)
----> 6 x_test_adv = attack.generate(x=scaledXTest)

/Library/Frameworks/Python.framework/Versions/3.9/lib/python3.9/site-packages/art/attacks/attack.py in replacement_function(self, *args, **kwargs)
73 if len(args) > 0:
74 args = tuple(lst)
---> 75 return fdict[func_name](self, *args, **kwargs)
76
77 replacement_function.doc = fdict[func_name].doc

/Library/Frameworks/Python.framework/Versions/3.9/lib/python3.9/site-packages/art/attacks/evasion/fast_gradient.py in generate(self, x, y, **kwargs)
221 # Use model predictions as correct outputs
222 logger.info("Using model predictions as correct labels for FGM.")
--> 223 y = get_labels_np_array(self.estimator.predict(x, batch_size=self.batch_size)) # type: ignore
224 y = y / np.sum(y, axis=1, keepdims=True)
225

/Library/Frameworks/Python.framework/Versions/3.9/lib/python3.9/site-packages/art/estimators/classification/classifier.py in replacement_function(self, *args, **kwargs)
69 if len(args) > 0:
70 args = tuple(lst)
---> 71 return fdict[func_name](self, *args, **kwargs)
72
73 replacement_function.doc = fdict[func_name].doc

/Library/Frameworks/Python.framework/Versions/3.9/lib/python3.9/site-packages/art/estimators/classification/scikitlearn.py in predict(self, x, **kwargs)
1507 y_pred_label = self.model.predict(X=x_preprocessed)
1508 targets = np.array(y_pred_label).reshape(-1)
-> 1509 one_hot_targets = np.eye(self.nb_classes)[targets]
1510 y_pred = one_hot_targets
1511

IndexError: arrays used as indices must be of integer (or boolean) type

[beat-buesser]

Hi @melikedem Thank you very much for using ART!
Could you please describe the shape of your input and label data? Is it binary or multi-class classification?

[melikedem]

Thank you so much for your fast reply!
It is multi-class classification.
X: (3550, 73)
Y: (3550, 1)

[beat-buesser]

Thank you! I have not yet been able to reproduce it, but I'm interested to see why this happens. Would you be able to share a code example that I can run to reproduces the error? It can work with random example data in case you are working on private data.

[melikedem]

Thank you again!
I attached my code. Yes I am working on my private data which is attached as well.
Datasets.zip
fast gradient.ipynb.zip

[beat-buesser]

I have noticed two items in your notebook:

• sklearn.svm.SVC.predict returns the same dtype as has been used in sklearn.svm.SVC.fit and in line svc.fit(scaledXTrain, y_train.values.ravel()) the ravel() returns indices of type np.float64. The current version of SklearnClassifier only supports labels of dtype=np.int*
• after fixing the first item, there is another issue, the labels in y_train are 1 to 14 for the 14 classes, but art.estimators.classification is expecting label indices starting from 0, therefore y_train should be 0 to 13

Changing svc.fit(scaledXTrain, y_train.values.ravel()) to svc.fit(scaledXTrain, y_train.values.ravel().astype(np.int)-1) is one way of fixing both, but storing the new labels in a separate variable would be better.

One item we learned from your question is that we should think about that model predictions in SklearnClassifier don't have to be always np.int and that we probably should add a check and conversion for other dtypes.

[melikedem]

Thank you so much for pointing out the problem.
If it is possible to support other data types with a simple check that would be great!
Have a nice day!