Skip to content

Latest commit

 

History

History
91 lines (57 loc) · 3.75 KB

index.md

File metadata and controls

91 lines (57 loc) · 3.75 KB
layout
home

Logo

This site introduces GloRo Nets (for "globally robust networks"), a family of Deep Neural Networks that is verifiably robust against L2-norm-bounded perturbations. Our Lipschitz-based robustness verification is instant and deterministic (so there is no false positive case) and scales favorably well to ImageNet-scale models. GloRo Nets have inspired a set of follow-up works in deep learning safety, explainability, overfitting and privacy. Read each individual post [HERE] for more information about GloRo-based works.

Contributions of GloRo Nets to Robustness Research

State-of-the-art Provable Robustness

(last update: 2023-06)

GloRo Nets provide the state-of-the-art deterministic robustness guarantee. We provide a quick overview of the best VRA (verified-robust accuracy) results here. These are more up-to-date and may surpass the results reported in the original paper.

dataset norm radius architecture VRA (%)
MNIST l2 1.58 Conv 4C3F 62.8
CIFAR-10 L2 0.141 LiResNet L12W512 70.1
CIFAR-100 L2 0.141 LiResNet L12W512 41.5
Tiny-Imagenet L2 0.141 LiResNet L12W512 33.6
ImageNet L2 0.141 LiResNet L12W588 35.0

📖 Read our ICML paper for GloRo Nets and the recent follow-up paper of LiResNet architecture.

💻 Check out our implementations for popular deep learning frameworks (click the icon below).

Towards Provable Top-K Robustness

TODO

Exploring Overfitting with GloRo Nets and TruLens

TODO

Exploring Privacy Leakage in Robust Models with GloRo Nets

TODO

A Pitfall of Robustness Certification: A Denial-of-service Attack.

TODO

Making A Pull Request for GloRo-based Projects and Publications

Please feel free to make a pull request at the github page of this website (https://github.com/cmu-transparency/gloronet) to include your own GloRo-based work. A pull request should add another MARKDOWN file to _posts/ following the template _templates/project_post.md.

Bibtex Citation

If you use the code of GloRo Net in your own project, please consider to using the following citations. If you are using a particular follow-up work described in [HERE], please additionally include the citation at the end of the project post.

@INPROCEEDINGS{leino21gloro,
    title = {Globally-Robust Neural Networks},
    author = {Klas Leino and Zifan Wang and Matt Fredrikson},
    booktitle = {International Conference on Machine Learning (ICML)},
    year = {2021}
}

@misc{kaiscaling2023,
  author = {Hu, Kai and Zou, Andy and Wang, Zifan and Leino, Klas and Fredrikson, Matt},
  title = {Scaling in Depth: Unlocking Robustness Certification on ImageNet},
  publisher = {arXiv},
  year = {2023}
}

Main Contributors

Klas Leino, Zifan Wang, Matt Fredrikson, Kai Hu, Andy Zou