TensorFlowV2Classifier and SaliencyMapMethod

[fatimah-aloraini]

Hello, thank you for ART. It is a great help!
I am working on adversarial attacks against IDS project (binary classification: normal and malicious samples); I faced "slice index 1 of dimension 1 out of bounds. [Op:StridedSlice] name: strided_slice/" error, and the traceback is attached. I did not understand why the line:
prediction=predictions[:unique_label]
in TensorFlowV2Classifier causes an error

`
from art.estimators.classification import TensorFlowV2Classifier
from art.attacks.evasion import SaliencyMapMethod
classifier = TensorFlowV2Classifier(
model=model,
loss_object=tf.keras.losses.BinaryCrossentropy(),
nb_classes=2,
input_shape=x_train[0].shape,
)

JSMA= SaliencyMapMethod(classifier=classifier, theta=0.1, gamma=0.5 )
x_test_adv = JSMA.generate(np.array(x_test))`

[beat-buesser]

Hi @fatimah-aloraini What is the shape of the model's pedictions? What is the accuracy of the model on its training data?

[fatimah-aloraini]

Hello @beat-buesser, appreciate your response.
The shape of the model prediction is TensorShape([219397, 1]) where 219397 is the number of samples in the testing set. The accuracy of the model on its training set is 96%

[beat-buesser]

Hi @fatimah-aloraini The attack uses class gradients which supports binary classification but requires a model with 2 output logits (1 for each class). Can you try with a model that has output shape TensorShape([219397, 2])?