FGSM for Sklearn MLPClassifier

[mccarthyajb]

Is your feature request related to a problem? Please describe.
*I am trying to implement FGSM with an MLPClassifier SklearnClassifier
Unfortunately I get the error message

art.exceptions.EstimatorError: FastGradientMethod requires an estimator derived from <class  'art.estimators.estimator.BaseEstimator'> and <class 'art.estimators.estimator.LossGradientsMixin'>, the provided classifier is an instance of <class 'art.estimators.classification.scikitlearn.ScikitlearnClassifier'> and is derived from (<class 'art.estimators.classification.classifier.ClassifierMixin'>, <class 'art.estimators.scikitlearn.ScikitlearnEstimator'>).

Describe the solution you'd like
I expect FGSM should work with Sklearn MLPClassifier. I wonder if MLPClassifier is currently supported?

Describe alternatives you've considered
I have considered building and using using a KerasClassifier instead, and had some success generating FGSM examples.

Additional context
Code Block

orig_model = MLPClassifier(
        hidden_layer_sizes=(256,128,64,32),
        activation="relu",
       random_state=42,verbose=True,
        max_iter=epochs)

orig_model.fit(x_train,y_train)

# Step 3: Create the ART classifier

art_classifier = SklearnClassifier(orig_model)
#art_classifier.fit(x_train,y_train)

# Step 6: Generate adversarial test examples



if(DO_FGSM):
    attack = FastGradientMethod(estimator=art_classifier, eps=0.5)
    print("Generating FGSM")
    x_test_adv = attack.generate(x=x_test)

[beat-buesser]

Hi @mccarthyajb Thank you very much for exploring ART! Yes, MLPClassifier is not yet supported, the reasons for this are likely that most users of ART would create their MLP models in one of the deep learning frameworks, a second reason is that to support MLPClassifier we would have to explicitly implement its gradient back-propagation function which the deep learning frameworks provide automatically for free. If you would be interested to implement support for MLPClassifier and its gradient calculation let me know, we would be happy to provide support.

To run your experiments, I would recommend creating the model in a deep learning framework. You have mentioned KerasClassifier, I would recommend its get-started example: https://github.com/Trusted-AI/adversarial-robustness-toolbox/blob/main/examples/get_started_keras.py