GiulioZizzo

Implementation of Certified Training based on IBP Methods

1

**Is your feature request related to a problem? Please describe.** Certified training for Abstract Interpretation/Interval bound propagation methods have been shown to boost the achievable certified performance. Adding these to...

**Is your feature request related to a problem? Please describe.** Denoised smoothing is a certified defence which splits the Randomised smoothing in two, firstly a pass through a _de-noising_ model,...

PGD attack on multi-modal CLIP model

2

# Description We include an experimental feature on handling multimodal inputs in ART and demonstrate its use for attacking CLIP. This is a new type of attack for ART and...

The current version of pylint on the CI pipeline (2.12.2) requires several ignore statements which are not needed by newer pylint versions (e.g. 2.17, 3.0.3, etc). For example, a common...

Currently, many ART estimators and tools do not have progress bars to indicate how quickly they are training, or to display useful information such as loss and accuracy. With ART...

Double Softmax in PyTorch image estimator for test cases.

2

Many tests use the PyTorch image estimator defined in the test utils. By default this estimator does _not_ use logits, e.g. the function signature is: `get_image_classifier_pt(from_logits=False, load_init=True, use_maxpool=True)` However, the...

Currently PyTorchClassifier determines the internal structure of the provided model via self._model._modules.items(). This is quite a brittle approach as it it purely dependant on how the layers were declared in...

Inclusion of Huggingface Support

5

Huggingface is one of the most popular libraries for carrying out machine learning on large pre-trained foundation models. With growing interest and activity in this area, we can aim to...