Baddet poisoning attack detector.fit()

[antilaanssi]

In this notebook: https://github.com/Trusted-AI/adversarial-robustness-toolbox/blob/main/notebooks/poisoning_attack_bad_det.ipynb

The notebook runs correctly until the very last line, ending in the following error:

---

TypeError Traceback (most recent call last)
Cell In[16], line 1
----> 1 detector.fit(x_poisoned, y_poisoned, nb_epochs=1)

File E:\sepate_tests_for_thesis\env\lib\site-packages\art\estimators\object_detection\pytorch_object_detector.py:425, in PyTorchObjectDetector.fit(self, x, y, batch_size, nb_epochs, drop_last, scheduler, **kwargs)
422 raise ValueError("An optimizer is needed to train the model, but none for provided.")
424 # Apply preprocessing and convert to tensors
--> 425 x_preprocessed, y_preprocessed = self._preprocess_and_convert_inputs(x=x, y=y, fit=True, no_grad=True)
427 class ObjectDetectionDataset(Dataset):
428 """
429 Object detection dataset in PyTorch.
430 """

File E:\sepate_tests_for_thesis\env\lib\site-packages\art\estimators\object_detection\pytorch_object_detector.py:217, in PyTorchObjectDetector._preprocess_and_convert_inputs(self, x, y, fit, no_grad)
214 x_tensor, y_tensor = cast_inputs_to_pt(x, y)
216 if not self.channels_first:
--> 217 x_tensor = torch.permute(x_tensor, (0, 3, 1, 2))
218 x_tensor = x_tensor / norm_factor
220 # Set gradients

TypeError: permute(): argument 'input' (position 1) must be Tensor, not list

I tried changing the input type to tensors and a numpy array, but it still fails.

Any ideas why this might be or workarounds for the issue?

[beat-buesser]

Hi @antilaanssi What is the type of x_poisoned in the last cell?

[antilaanssi]

It is a list of numpy arrays. It should be a numpy array of numpy arrays for detector.fit() to work. However making the list of numpy arrays to numpy array of numpy arrays gives the following error:

---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
/run/nvme/job_20432149/tmp/ipykernel_2099131/1850239494.py in <cell line: 4>()
      2 
      3 print(type(x_poisoned[0]))
----> 4 np_array_of_arrays = np.array(x_poisoned)

ValueError: setting an array element with a sequence. The requested array has an inhomogeneous shape after 1 dimensions. The detected shape was (2,) + inhomogeneous part.

If calling detector.fit() with just one numpy.ndarray from the list, I get the following error:

---------------------------------------------------------------------------
RuntimeError                              Traceback (most recent call last)
/run/nvme/job_20432149/tmp/ipykernel_2099131/2349275260.py in <cell line: 1>()
----> 1 detector.fit(x_poisoned[0], y_poisoned, nb_epochs=1)

/scratch/project_2008539/lib/python3.10/site-packages/art/estimators/object_detection/pytorch_object_detector.py in fit(self, x, y, batch_size, nb_epochs, drop_last, scheduler, **kwargs)
    423 
    424         # Apply preprocessing and convert to tensors
--> 425         x_preprocessed, y_preprocessed = self._preprocess_and_convert_inputs(x=x, y=y, fit=True, no_grad=True)
    426 
    427         class ObjectDetectionDataset(Dataset):

/scratch/project_2008539/lib/python3.10/site-packages/art/estimators/object_detection/pytorch_object_detector.py in _preprocess_and_convert_inputs(self, x, y, fit, no_grad)
    215 
    216             if not self.channels_first:
--> 217                 x_tensor = torch.permute(x_tensor, (0, 3, 1, 2))
    218             x_tensor = x_tensor / norm_factor
    219 

RuntimeError: number of dims don't match in permute

Any ideas?