ug100 1.0.0

Dataset containing adversarial results for seven approximate attacks (+ MIP) on MNIST and CIFAR10.

UG100

A dataset containing adversarial results for seven approximate attacks (+ MIP) on a subset of the MNIST and CIFAR10 test datasets. Specifically, it contains ~2.3k adversarial examples generated by the following attacks:

• Basic Iterative Method (bim)
• Brendel & Bethge Attack (brendel)
• Carlini & Wagner Attack (carlini)
• Deepfool (deepfool)
• Fast Gradient Sign Method (fast_gradient)
• Projected Gradient Descent (pgd)
• Uniform noise (uniform)
• MIPVerify (mip)

It also includes adversarial distances (for all attacks) and bounds (for MIP), as well as MIP convergence times.

Applications of this dataset include:

• Studying how, when and why adversarial attacks are close-to-optimal;
• Training classifiers that are robust to adversarial noise;
• Benchmarking new adversarial attacks.

The code used to generate UG100 can be found here.

Installation

pip install ug100

Implementation Notes

Since there aren't adversarial examples for every element of the test sets, we store the adversarials as an index-to-results dictionary. For sequential access, use IndexDataset.

Additionally, we do not store the corresponding genuine examples for MNIST and CIFAR10. If you're using PyTorch, consider using TorchVision's dataset library.

Citing this Dataset

Please cite this dataset as:

Samuele Marro and Michele Lombardi. _Asymmetries in Adversarial Settings_. 2022.