Error in Zoo attack about adversarial-robustness-toolbox HOT 9 CLOSED

Fixed in b0ecc55.

from adversarial-robustness-toolbox.

Bug still there (ART version 1.3.1)

Traceback (most recent call last):
  File "generate_adv.py", line 194, in <module>
    main(args)
  File "generate_adv.py", line 175, in main
    adv_data = attack.generate(x=x_test, y=y_test)
  File "/home/aaldahdo/adv_dnn/venv/lib/python3.8/site-packages/art/attacks/attack.py", line 74, in replacement_function
    return fdict[func_name](self, *args, **kwargs)
  File "/home/aaldahdo/adv_dnn/venv/lib/python3.8/site-packages/art/attacks/evasion/zoo.py", line 222, in generate
    res = self._generate_batch(x_batch, y_batch)
  File "/home/aaldahdo/adv_dnn/venv/lib/python3.8/site-packages/art/attacks/evasion/zoo.py", line 263, in _generate_batch
    best_dist, best_label, best_attack = self._generate_bss(x_batch, y_batch, c_current)
  File "/home/aaldahdo/adv_dnn/venv/lib/python3.8/site-packages/art/attacks/evasion/zoo.py", line 380, in _generate_bss
    x_adv = self._optimizer(x_adv, y_batch, c_batch)
  File "/home/aaldahdo/adv_dnn/venv/lib/python3.8/site-packages/art/attacks/evasion/zoo.py", line 455, in _optimizer
    expanded_x, expanded_x + coord_batch.reshape(expanded_x.shape), expanded_targets, expanded_c,
ValueError: cannot reshape array of size 25165824 into shape (4096,32,32,3)

from adversarial-robustness-toolbox.

Hi @aldahdooh, thank you for your message. I was not able to reproduce your observation with branch dev_1.4.3 based on example get_started_pytorch.pychange to useZooAttack`. Could you please provide more details or code example how to reproduce the error?

from adversarial-robustness-toolbox.

For CIFAR-10 dataset with keras model
Note: No error with MNIST

attack = ZooAttack(classifier=kclassifier, batch_size=32)
adv_data = attack.generate(x=x_test, y=y_test)

from adversarial-robustness-toolbox.

Hi @aldahdooh I have tried to reproduce this issue with dev_1.3.1 and dev_1.4.3 using get_started_keras.py and have not been able to reproduce the issue with the attack definition above for the first 32 test samples. Would you be able to show the entire code example producing this issue?

from adversarial-robustness-toolbox.

zoo_test.zip

Thanks for your reply, Attached is the code. Just run generate_adv.py -d=cifar. I run it with MNIST and no errors encountered. With CIFAR, the problem came with the last batch iteration.

from adversarial-robustness-toolbox.

Oh, I see, it happens in the last batch iteration, thank you, that's very helpful. Let me try again using your code.

from adversarial-robustness-toolbox.

Hi @aldahdooh I have pushed an update for ZOO to #755 that should fix the issue. If you have time, could you please try it out and let me know if it works for you?

I have noticed that you are creating the attacks in generate_adv.py with the attack's default arguments which are selected to let the attack run fast but not necessarily very strong.

from adversarial-robustness-toolbox.

Hi @beat-buesser, Many thanks for your feedback. I will do that ASAP since my machine is busy now with other tasks.
And thanks for your note, I do agree with you, I did that to compare with other works that use same settings.

from adversarial-robustness-toolbox.