Use targeted carlini in UniversalPerturbation to generate noise but fail.

[MarcuXu]

Describe the bug
I hope to use the targeted carlini in UniversalPerturbation to generate noise, but the noise I get is just zeros arrays. Could you please help me with that?

To Reproduce
Steps to reproduce the behavior:

1. Clone the 1.14 ART and use the UniversalPerturbation function in my codes.
2. Set related parameters like this:
   # build targeted UAP
   classifier = KerasClassifier(model=model)
   # print(type(model))
   adv_crafter = UniversalPerturbation(
   classifier,
   attacker='carlini',
   delta=1e-2,
   attacker_params={'verbose': False, 'batch_size': 64,
   'targeted': True,
   'confidence': 5e-2, 'learning_rate': 5e-2, 'max_iter': 10, 'initial_const': 5e-2},
   max_iter=20,
   verbose=True,
   # eps=5.5,
   # Only 1, 2 , np.Inf 1.1 and 1.2 supported for now.
   norm=np.Inf,)
   # set target label
   target = 3
   y_train_adv_tar = np.zeros(Y_train.shape)
   for i in range(Y_train.shape[0]):
   y_train_adv_tar[i, target] = 1
   # generate noise
   _ = adv_crafter.generate(X_train, y=y_train_adv_tar)
   noise = adv_crafter.noise
   print(noise)
   # The upper codes work well when I use FGSM in TargetedUniversalPerturbation, so I think the logic of my codes could be right. Thus, I may forget or don't know how to generate the noise under carlini correctly.
   Expected behavior
   Get some noise which can be used to attack.

Screenshots
If applicable, add screenshots to help explain your problem.

System information (please complete the following information):

• Linux
• Python version: 3.9
• ART version or commit number: 1.14.0
• TensorFlow / Keras / PyTorch / MXNet version: pip install . under the cloned ART.

[beat-buesser]

Hi @MarcuXu What is the output of your model, probabilities or logits?

I'll convert this issue into a discussion under the Discussions tab above to continue the discussion.

[MarcuXu]

Hi @beat-buesser, thank you for your reply! The inputs of training datasets are like x_train (1000*200*52) and y_train (1000*6), where 1000 is the sample number, 200*52 is the sample size, and 6 is the class number. Through the prediction from the CNN model (whose type is <class 'tensorflow.python.keras.engine.sequential.Sequential'>), like:

y = model.predict(x)
print(y.shape)
print(y[0])

# output:
# (200, 6)
# [9.9835736e-01 1.6425920e-03 6.0162335e-13 1.3994645e-16 1.0540235e-18 7.7685725e-09] 
# It looks like a one-hot array
#######

Could you please tell me how to implement the CW correctly in my CNN based on ART? I have been stuck in the problem for a week...