Closed mmirman closed 5 years ago
Sure, we'll do that
Awesome! Ping us when you're done, and we'll add the defense to the list.
Bump -- do you have any questions or issues with implementing the interface?
On pause until we update our repo with some larger changes
Bump - any update on this?
Name: Differentiable Abstract Interpretation for Provably Robust Neural Networks
Authors: Matthew Mirman, Timon Gehr, Martin Vechev
Paper: https://www.sri.inf.ethz.ch/papers/icml18-diffai.pdf
Code: https://github.com/eth-sri/diffai
Venue: ICML 2018
Does the code implement the robust-ml API and include pre-trained models: no
Dataset: MNIST
Threat model: $l_\infty(\epsilon = 0.1)$
Natural accuracy: 99%
Claims: 96.4% proved