Submitting a new defense (preprint)

[YupingLin171]

Name: {Bandlimiting Neural Networks Against Adversarial Attacks}

Authors: {Yuping Lin, Kasra Ahmadi K. A., Hui Jiang}

Paper: {https://arxiv.org/abs/1905.12797}

Code: {https://github.com/YupingLin171/PostAvgDefense}

Venue: {N/A}

Does the code implement the robust-ml API and include pre-trained models: {yes}

Dataset: {ImageNet, CIFAR-10}

Threat model: {ℓ∞(ϵ=8/255)}

Natural accuracy: {77.32% on ImageNet, 92.55% on CIFAR-10}

Claims: {76.06% on ImageNet, 88.41% on CIFAR-10}

Note: {These are the new results and haven't updated to the paper in arXiv yet. The code is up to date. We have tested the robust-ml API implementation on CIFAR-10 but haven't tested on ImageNet (didn't find the "val.txt" file the data provider needs). We just assumed data provider provides data in (N, H, W, C) arrangement.}

[anishathalye]

Thank you for your submission! We've added it here: https://www.robust-ml.org/preprints/