A Kernelized Manifold Mapping to Diminish the Effect of Adversarial Perturbations

[asgsaeid]

Name: {A Kernelized Manifold Mapping to Diminish the Effect of Adversarial Perturbations}

Authors: {Saeid Asgari Taghanaki, Kumar Abhishek, Shekoofeh Azizi, Ghassan Hamarneh}

Paper: {https://arxiv.org/abs/1903.01015}

Code: {https://github.com/asgsaeid/KernelizedManifoldMapping}

Venue: {CVPR 2019}

Does the code implement the robust-ml API and include pre-trained models: {no}

Dataset: {MNIST, ISIC skin, chest x-ray}

Threat model: {PGD, e=0.1}

Natural accuracy: {99.42 (MNIST)}

Claims: {81.57 (MNIST, PGD)}

[anishathalye]

Thank you for your submission!

• We currently don't support the ISIC skin or chest x-ray dataset. If you would like, you can contribute code to support this (e.g. as in https://github.com/robust-ml/robust-ml.github.io/issues/5). In the meantime, if you would like, we can list your submission with the MNIST claim only.
• Per our policy, your code must implement the robust-ml API and include pre-trained models. It should be minimal effort to do this; you can see previous submissions to see what kind of code needed to be added for this.
• What is the metric for the threat model? I'm guessing L-infinity, but it would be good if you can confirm.
• Claims are "accuracy under a threat model", and they cannot restrict the attacker to a particular type of attack (e.g. PGD), so that claim would just be listed as "81.57% accuracy".

[anishathalye]

Closing due to inactivity. If you do implement the Robust-ML API, please re-open this issue and we can add the defense to the list.